Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed because of a cast error
Error code: DatasetGenerationCastError
Exception: DatasetGenerationCastError
Message: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 6 new columns ({'completion_by_task', 'task_completion_summary', 'workflow_analysis', 'total_videos', 'export_timestamp', 'quality_metrics'}) and 3 missing columns ({'video_id', 'captions', 'video_url'}).
This happened while the json dataset builder was generating data using
hf://datasets/zhiqiulin/caption_export/export_20250915_2141/comprehensive_statistics_20250915_2141.json (at revision 4a996b0920c2fb1b3c4772397ab2d68c08090881)
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1831, in _prepare_split_single
writer.write_table(table)
File "/usr/local/lib/python3.12/site-packages/datasets/arrow_writer.py", line 714, in write_table
pa_table = table_cast(pa_table, self._schema)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/table.py", line 2272, in table_cast
return cast_table_to_schema(table, schema)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/table.py", line 2218, in cast_table_to_schema
raise CastError(
datasets.table.CastError: Couldn't cast
export_timestamp: string
total_videos: int64
task_completion_summary: struct<total_possible_tasks: int64, completed_tasks: int64, reviewed_tasks: int64, completion_rate: (... 28 chars omitted)
child 0, total_possible_tasks: int64
child 1, completed_tasks: int64
child 2, reviewed_tasks: int64
child 3, completion_rate: double
child 4, review_rate: double
workflow_analysis: struct<improved_precaption: int64, perfect_precaption: int64>
child 0, improved_precaption: int64
child 1, perfect_precaption: int64
quality_metrics: struct<initial_ratings: list<item: int64>, final_ratings: list<item: int64>, avg_initial_rating: dou (... 30 chars omitted)
child 0, initial_ratings: list<item: int64>
child 0, item: int64
child 1, final_ratings: list<item: int64>
child 0, item: int64
child 2, avg_initial_rating: double
child 3, avg_final_rating: double
completion_by_task: struct<subject: struct<approved: int64, rejected: int64, completed_not_reviewed: int64, not_complete (... 478 chars omitted)
child 0, subject: struct<approved: int64, rejected: int64, completed_not_reviewed: int64, not_completed: int64>
child 0, approved: int64
child 1, rejected: int64
child 2, completed_not_reviewed: int64
child 3, not_completed: int64
child 1, scene: struct<approved: int64, rejected: int64, completed_not_reviewed: int64, not_completed: int64>
child 0, approved: int64
child 1, rejected: int64
child 2, completed_not_reviewed: int64
child 3, not_completed: int64
child 2, motion: struct<rejected: int64, approved: int64, completed_not_reviewed: int64, not_completed: int64>
child 0, rejected: int64
child 1, approved: int64
child 2, completed_not_reviewed: int64
child 3, not_completed: int64
child 3, spatial: struct<approved: int64, rejected: int64, completed_not_reviewed: int64, not_completed: int64>
child 0, approved: int64
child 1, rejected: int64
child 2, completed_not_reviewed: int64
child 3, not_completed: int64
child 4, camera: struct<rejected: int64, approved: int64, completed_not_reviewed: int64, not_completed: int64>
child 0, rejected: int64
child 1, approved: int64
child 2, completed_not_reviewed: int64
child 3, not_completed: int64
child 5, color: struct<>
child 6, lighting: struct<>
child 7, effects: struct<>
to
{'video_id': Value('string'), 'video_url': Value('string'), 'captions': {'camera': {'caption_data': {'caption_rating_score': Value('int64'), 'feedback_is_needed': Value('bool'), 'feedback_rating_score': Value('int64'), 'final_caption': Value('string'), 'final_feedback': Value('string'), 'gpt_caption': Value('string'), 'gpt_caption_llm': Value('string'), 'gpt_caption_prompt': Value('string'), 'gpt_feedback': Value('string'), 'gpt_feedback_llm': Value('string'), 'gpt_feedback_prompt': Value('string'), 'initial_caption_rating_score': Value('int64'), 'initial_feedback': Value('string'), 'original_annotator': Value('string'), 'pre_caption': Value('string'), 'pre_caption_llm': Value('string'), 'pre_caption_mode': Value('string'), 'pre_caption_prompt': Value('string'), 'review_timestamp': Value('string'), 'reviewer': Value('string'), 'reviewer_double_check': Value('bool'), 'timestamp': Value('string'), 'user': Value('string'), 'workflow_type': Value('string')}, 'negative_example': {'caption_rating_score': Value('int64'), 'feedback_is_needed': Value('bool'), 'feedback_rating_score': Value('int64'), 'final_caption': Value('string'), 'final_feedback': Value('string'), 'gpt_caption': Value('string'), 'gpt_caption_llm': Value('string'), 'gpt_caption_prompt': Value('string'), 'gpt_feedback': Value('string'), 'gpt_feedback_llm': Value('string'), 'gpt_feedback_prompt': Value('string'), 'initial_caption_rating_score': Value('int64'), 'initial_feedback': Value('string'), 'pre_caption': Value(
...
('string'), 'final_feedback': Value('string'), 'gpt_caption': Value('string'), 'gpt_caption_llm': Value('string'), 'gpt_caption_prompt': Value('string'), 'gpt_feedback': Value('string'), 'gpt_feedback_llm': Value('string'), 'gpt_feedback_prompt': Value('string'), 'initial_caption_rating_score': Value('int64'), 'initial_feedback': Value('string'), 'original_annotator': Value('string'), 'pre_caption': Value('string'), 'pre_caption_llm': Value('string'), 'pre_caption_mode': Value('string'), 'pre_caption_prompt': Value('string'), 'review_timestamp': Value('string'), 'reviewer': Value('string'), 'reviewer_double_check': Value('bool'), 'timestamp': Value('string'), 'user': Value('string'), 'workflow_type': Value('string')}, 'negative_example': {'caption_rating_score': Value('int64'), 'feedback_is_needed': Value('bool'), 'feedback_rating_score': Value('int64'), 'final_caption': Value('string'), 'final_feedback': Value('string'), 'gpt_caption': Value('string'), 'gpt_caption_llm': Value('string'), 'gpt_caption_prompt': Value('string'), 'gpt_feedback': Value('string'), 'gpt_feedback_llm': Value('string'), 'gpt_feedback_prompt': Value('string'), 'initial_caption_rating_score': Value('int64'), 'initial_feedback': Value('string'), 'pre_caption': Value('string'), 'pre_caption_llm': Value('string'), 'pre_caption_mode': Value('string'), 'pre_caption_prompt': Value('string'), 'timestamp': Value('string'), 'user': Value('string'), 'workflow_type': Value('string')}, 'status': Value('string')}}}
because column names don't match
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1455, in compute_config_parquet_and_info_response
parquet_operations = convert_to_parquet(builder)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1054, in convert_to_parquet
builder.download_and_prepare(
File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 894, in download_and_prepare
self._download_and_prepare(
File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 970, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1702, in _prepare_split
for job_id, done, content in self._prepare_split_single(
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1833, in _prepare_split_single
raise DatasetGenerationCastError.from_cast_error(
datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 6 new columns ({'completion_by_task', 'task_completion_summary', 'workflow_analysis', 'total_videos', 'export_timestamp', 'quality_metrics'}) and 3 missing columns ({'video_id', 'captions', 'video_url'}).
This happened while the json dataset builder was generating data using
hf://datasets/zhiqiulin/caption_export/export_20250915_2141/comprehensive_statistics_20250915_2141.json (at revision 4a996b0920c2fb1b3c4772397ab2d68c08090881)
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
video_id
string | video_url
string | captions
dict |
|---|---|---|
-2uIa-XMJC0.3.0.mp4
|
https://huggingface.co/datasets/zhiqiulin/video_captioning/resolve/main/-2uIa-XMJC0.3.0.mp4
|
{
"camera": {
"caption_data": {
"caption_rating_score": 5,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The camera is positioned at eye level and begins with a level angle, looking straight ahead. The camera uses deep focus, maintaining clarity across the scene. It slightly pans right to center the skateboarder and then moves forward, subtly tilting upward while maintaining a level angle and focusing straight ahead. The movement is slightly unsteady, with some shaking.",
"final_feedback": "As the camera advances toward the skateboarder, it subtly tilts upward while maintaining a level angle and continues to focus straight ahead.",
"gpt_caption": "The camera is positioned at eye level and begins with a level angle, looking straight ahead. It then shifts to a low angle, looking up at the skateboarder. The camera uses deep focus, maintaining clarity across the scene. It slightly pans right to center the skateboarder and then moves forward, subtly tilting upward while maintaining a level angle and focusing straight ahead. The movement is slightly unsteady, with some shaking.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "As the camera advances toward the skateboarder, it subtly tilts upward while maintaining a level angle and continues to focus straight ahead.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 4,
"initial_feedback": "As the camera moves toward the skateboarder, it slightly tilts upward but maintains a level angle and continues to look straight ahead",
"original_annotator": "Yuhan Huang",
"pre_caption": "The camera is positioned at eye level and begins with a level angle, looking straight ahead. It then shifts to a low angle, looking up at the skateboarder. The camera uses deep focus, maintaining clarity across the scene. It slightly pans right to center the skateboarder and then moves forward to approach them. The movement is slightly unsteady, with some shaking.",
"pre_caption_llm": "gemini-2.5-pro-preview-05-06",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the **video’s and camera’s configuration**, covering **video speed, lens distortion, camera angle, camera height, movements (translation, rotation, steadiness, intensity, complexity), and focus (depth, focus plane, focus changes).** \n\nIf **video speed** is altered, specify the type, such as *time-lapse* (“Clouds move rapidly across the sky”), *slow-motion*, or *speed ramp* (a mix of fast and slow speeds). If the video is *time-reversed* or *stop-motion*, note this as well. \n\nIf **lens distortion** is present, describe the type and degree. For example, *fisheye distortion* creates extreme curvature, while *barrel distortion* causes mild outward bowing of straight lines near the edges. \n\nDescribe the **camera height** in relation to the ground, such as *eye-level, ground-level, overhead-level, or underwater*. If height changes due to movement, mention how it transitions. Similarly, specify the **camera angle**, such as *bird’s eye, high angle, low angle, or worm’s eye*, noting any shifts within the video. If a **Dutch angle** (tilted horizon) is present, indicate whether it remains fixed or varies due to camera rolling. \n\nIf discernible, describe the **camera focus and depth of field**. For example, *deep focus* keeps all elements sharp, while *shallow depth of field* blurs the background or foreground. If focus changes dynamically, note whether it’s a *rack focus* (shifting focus between subjects) or *focus tracking* (following a subject’s depth movement). If the video lacks realistic depth of field, describe whether it appears artificial, overly blurry, or lacking in cinematic focus effects. \n\nIf the **camera is static**, no further description is needed. If it moves, describe its **type, direction, and speed**. Specify movements such as *tracking (following a subject), arcing (circling around), dollying (moving forward/backward), trucking (moving left/right), panning (rotating horizontally), tilting (angling vertically), pedestaling (lifting/lowering), or rolling (rotating around the lens axis).* If motion intensity varies, distinguish between slow and rapid movements and present them in **temporal order** (e.g., *“The camera first pans right, then tilts upward to follow the subject”*). If motion appears **fragmented or random**, summarize it concisely rather than over-describing each small movement (e.g., *“The first-person camera moves erratically, frequently changing direction without a clear pattern”*).\n\nCrucially, instead of inferring these attributes from the video, we have already provided human-labeled ground truth for some of the elements specified above. You should directly use this information in your description and should not infer any details that are not already provided. Your description should be brief, and if anything is normal or unremarkable, you do not need to include it (e.g., if the video is at regular playback speed, there is no need to mention it).\n\n**Playback Speed:** The playback is forward at real-time speed without noticeable speed-up or slowdown. (no need to mention).\n\n**Lens Distortion:** No lens distortion (no need to mention).\n\n**Camera Height:** The camera is at an eye level (above the waist).\n\n**Camera Angle:** The camera angle is initially at a level angle (looking straight ahead) and then changes to a low angle (looking up from below) due to camera motion.\n\n**Camera Focus:** The camera uses a deep focus with a large depth of field.\n\n**Camera Motion:** The camera slightly pans right to center the skateboarder, then moves forward to approach it.\n\n**Camera Steadiness:** The camera movement is slightly unsteady with some shaking.\n\nIf possible, specify the subject that the camera focuses on when describing camera work. For instance, use 'focus on the man in the foreground' rather than 'focus on the foreground.' Likewise, if the camera follows a subject, avoid the generic phrase 'tracking the subject(s).' Instead, identify the subject and describe the specific type of tracking.",
"review_timestamp": "2025-05-29T02:17:40.824495",
"reviewer": "Shihang Zhu",
"reviewer_double_check": false,
"timestamp": "2025-05-29T02:34:07.348052",
"user": "Shihang Zhu",
"workflow_type": "improved_precaption"
},
"negative_example": {
"caption_rating_score": null,
"feedback_is_needed": false,
"feedback_rating_score": null,
"final_caption": "The camera is positioned at eye level and begins with a level angle, looking straight ahead. It then shifts to a low angle, looking up at the skateboarder. The camera uses deep focus, maintaining clarity across the scene. It slightly pans right to center the skateboarder and then moves forward to approach them. The movement is slightly unsteady, with some shaking.",
"final_feedback": null,
"gpt_caption": null,
"gpt_caption_llm": null,
"gpt_caption_prompt": null,
"gpt_feedback": null,
"gpt_feedback_llm": null,
"gpt_feedback_prompt": null,
"initial_caption_rating_score": 5,
"initial_feedback": null,
"pre_caption": "The camera is positioned at eye level and begins with a level angle, looking straight ahead. It then shifts to a low angle, looking up at the skateboarder. The camera uses deep focus, maintaining clarity across the scene. It slightly pans right to center the skateboarder and then moves forward to approach them. The movement is slightly unsteady, with some shaking.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the **video’s and camera’s configuration**, covering **video speed, lens distortion, camera angle, camera height, movements (translation, rotation, steadiness, intensity, complexity), and focus (depth, focus plane, focus changes).** \n\nIf **video speed** is altered, specify the type, such as *time-lapse* (“Clouds move rapidly across the sky”), *slow-motion*, or *speed ramp* (a mix of fast and slow speeds). If the video is *time-reversed* or *stop-motion*, note this as well. \n\nIf **lens distortion** is present, describe the type and degree. For example, *fisheye distortion* creates extreme curvature, while *barrel distortion* causes mild outward bowing of straight lines near the edges. \n\nDescribe the **camera height** in relation to the ground, such as *eye-level, ground-level, overhead-level, or underwater*. If height changes due to movement, mention how it transitions. Similarly, specify the **camera angle**, such as *bird’s eye, high angle, low angle, or worm’s eye*, noting any shifts within the video. If a **Dutch angle** (tilted horizon) is present, indicate whether it remains fixed or varies due to camera rolling. \n\nIf discernible, describe the **camera focus and depth of field**. For example, *deep focus* keeps all elements sharp, while *shallow depth of field* blurs the background or foreground. If focus changes dynamically, note whether it’s a *rack focus* (shifting focus between subjects) or *focus tracking* (following a subject’s depth movement). If the video lacks realistic depth of field, describe whether it appears artificial, overly blurry, or lacking in cinematic focus effects. \n\nIf the **camera is static**, no further description is needed. If it moves, describe its **type, direction, and speed**. Specify movements such as *tracking (following a subject), arcing (circling around), dollying (moving forward/backward), trucking (moving left/right), panning (rotating horizontally), tilting (angling vertically), pedestaling (lifting/lowering), or rolling (rotating around the lens axis).* If motion intensity varies, distinguish between slow and rapid movements and present them in **temporal order** (e.g., *“The camera first pans right, then tilts upward to follow the subject”*). If motion appears **fragmented or random**, summarize it concisely rather than over-describing each small movement (e.g., *“The first-person camera moves erratically, frequently changing direction without a clear pattern”*).\n\nCrucially, instead of inferring these attributes from the video, we have already provided human-labeled ground truth for some of the elements specified above. You should directly use this information in your description and should not infer any details that are not already provided. Your description should be brief, and if anything is normal or unremarkable, you do not need to include it (e.g., if the video is at regular playback speed, there is no need to mention it).\n\n**Playback Speed:** A regular-speed video where events play in real-time without acceleration or slowing. (no need to mention).\n\n**Lens Distortion:** No lens distortion (no need to mention).\n\n**Camera Height:** The camera is at an eye level (above the waist).\n\n**Camera Angle:** The camera angle is initially at a level angle (looking straight ahead) and then changes to a low angle (looking up from below) due to camera motion.\n\n**Camera Focus:** The camera uses a deep focus with a large depth of field.\n\n**Camera Motion:** The camera slightly pans right to center the skateboarder, then moves forward to approach it.\n\n**Camera Steadiness:** The camera movement is slightly unsteady with some shaking.\n\nIf possible, specify the subject that the camera focuses on when describing camera work. For instance, use 'focus on the man in the foreground' rather than 'focus on the foreground.' Likewise, if the camera follows a subject, avoid the generic phrase 'tracking the subject(s).' Instead, identify the subject and describe the specific type of tracking.",
"timestamp": "2025-03-30T10:18:06.331136",
"user": "Yuhan Huang",
"workflow_type": "perfect_precaption"
},
"status": "rejected"
},
"motion": {
"caption_data": {
"caption_rating_score": 5,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The video begins with a view of an indoor skate park, featuring a wooden ramp in the foreground and various skateboards and equipment in the background. The camera gradually zooms in on a woman standing on the ramp, who is wearing a white hat and light-colored clothing. As the camera gets closer, the woman remains stationary, looking forward. The background includes a neon sign that reads 'BRING ON THE NIGHT' and several skateboards mounted on the wall. Towards the end of the video, a drone enters the frame from the left, prompting the woman to turn her head to her right to watch it.",
"final_feedback": "Towards the end of the video, a drone enters the frame from the left. In response, the woman turns her head to her right to watch the drone.",
"gpt_caption": "The video begins with a view of an indoor skate park, featuring a wooden ramp in the foreground and various skateboards and equipment in the background. The camera gradually zooms in on a woman standing on the ramp, who is wearing a white hat and light-colored clothing. As the camera gets closer, the woman remains stationary, looking forward. The background includes a neon sign that reads 'BRING ON THE NIGHT' and several skateboards mounted on the wall. Towards the end of the video, a drone enters the frame from the left, prompting the woman to turn her head to the right to watch it.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "Towards the end of the video, a drone enters the frame from the left. In response, the woman turns her head to the left to watch the drone.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "At the conclusion of the video, a drone appears from the left side of the frame. Subsequently, the woman turns her head to the left of the frame to observe the drone.",
"original_annotator": "Yuhan Huang",
"pre_caption": "The video begins with a view of an indoor skate park, featuring a wooden ramp in the foreground and various skateboards and equipment in the background. The camera gradually zooms in on a woman standing on the ramp, who is wearing a white hat and light-colored clothing. As the camera gets closer, the woman remains stationary, looking forward. The background includes a neon sign that reads 'BRING ON THE NIGHT' and several skateboards mounted on the wall. The woman continues to stand on the ramp, maintaining her position as the camera focuses on her.",
"pre_caption_llm": "tarsier2-7b",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the subject’s motion in this video, ensuring actions are presented in **chronological order** if multiple movements occur (e.g., \"The bird first takes flight, then soars in a circle, and finally lands on a branch\"). Focus on the subject's motion rather than repeating details already included in the human-written subject descriptions. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B.' Return only the one-paragraph video description without Markdown formatting or introductory text.\n\nIf the subject in the video has no movement, please briefly mention that without going into too much detail.\n\nPlease only describe the content of the video. Don't mention the details of the subject's appearance unless you need to differentiate between multiple subjects by their appearance. Clearly describe the subject's motion.\n\nAvoid abstract descriptions, such as \"The car maintains a low, sleek profile as it maneuvers the bend, emphasizing its speed and agility\" and \"emphasizing its speed and agility as it maneuvers through the turn.\"\n\nBelow are detailed instructions:\n\nDescribe **individual subject actions** with clarity, specifying how they move rather than using generic descriptions. For example, instead of “a person is running,” say “a runner sprints across the finish line.”\n\nIf the subject interacts with an **object**, specify the type of interaction and its effect. Instead of “a person is working,” say “a construction worker operates a jackhammer, breaking up the pavement.”\n\nIf there are **interactions between subjects**, describe the nature of their relationship and movements relative to each other. Instead of “people are fighting,” say “two boxers exchange blows in the ring, circling each other cautiously.”\n\nIf there are collective behaviors for a group of subjects, describe **group activities** with specificity. Instead of “birds are flying,” say “a flock of geese flies in a V-formation across the horizon.” Instead of “people are walking,” say “a crowd of protesters marches down the street, carrying signs and banners.” Clearly convey the type of group, their coordinated actions, and any notable patterns in their movement.\n\nPlease note that the video features salient **human** subjects, so the description should focus on their motion and dynamics.\n\nIn addition, the human-written caption below already describes the subjects (if any) in this video but does not comprehensively capture their motion. Use this caption as a reference or starting point to draft the description, but do not rely on it completely. Expand or refine it to fully capture the subjects' motion and dynamics. However, if the caption does not mention any subjects, do not add any description—simply note this briefly in the description. Please note that this caption is only to help you clarify which subject's motion needs to be analyzed, rather than adding more content based on this description.\n\nHuman-Written Caption: **The video features a woman standing confidently on a wooden ramp in an indoor setting. She has short, light-colored hair and is wearing a light-colored knit hat, tank top, and pants, suggesting a casual style. The woman appears focused and poised, with a relaxed posture. Surrounding her are various skateboards mounted on the wall, adding a dynamic and sporty atmosphere to the scene. The lighting casts distinct shadows, enhancing the visual depth and highlighting her presence in the space. At the end of the video, a drone enters the frame from the left side.**",
"review_timestamp": "2025-05-27T03:36:55.734239",
"reviewer": "Shihang Zhu",
"reviewer_double_check": false,
"timestamp": "2025-05-27T03:39:17.116291",
"user": "Shihang Zhu",
"workflow_type": "improved_precaption"
},
"negative_example": {
"caption_rating_score": 4,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The video begins with a view of an indoor skate park, featuring a wooden ramp and various skateboards and equipment. The camera gradually zooms in on a woman standing on the ramp, who is wearing a white hat and light-colored clothing. As the camera gets closer, the woman remains stationary, looking forward. The background includes a neon sign that reads 'BRING ON THE NIGHT' and several skateboards mounted on the wall. Toward the end of the video, a drone enters the frame from the left, prompting the woman to turn her head to the left to look at it, while maintaining her position on the ramp.",
"final_feedback": "At the end of the video, a drone enters the frame from the left, prompting the woman to turn her head to the left to look at it.",
"gpt_caption": "The video begins with a view of an indoor skate park, featuring a wooden ramp in the foreground and various skateboards and equipment in the background. The camera gradually zooms in on a woman standing on the ramp, who is wearing a white hat and light-colored clothing. As the camera gets closer, the woman remains stationary, looking forward. The background includes a neon sign that reads 'BRING ON THE NIGHT' and several skateboards mounted on the wall. Toward the end of the video, a drone enters the frame from the left, prompting the woman to turn her head to the left to look at it, while maintaining her position on the ramp.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "At the end of the video, a drone enters the frame from the left, prompting the woman to turn her head to the left to look at it.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "视频最后一个无人机从左边飞进画面，然后人物头转向左边看着无人机",
"pre_caption": "The video begins with a view of an indoor skate park, featuring a wooden ramp in the foreground and various skateboards and equipment in the background. The camera gradually zooms in on a woman standing on the ramp, who is wearing a white hat and light-colored clothing. As the camera gets closer, the woman remains stationary, looking forward. The background includes a neon sign that reads 'BRING ON THE NIGHT' and several skateboards mounted on the wall. The woman continues to stand on the ramp, maintaining her position as the camera focuses on her.",
"pre_caption_llm": "tarsier-recap-7b",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the subject’s motion in this video, ensuring actions are presented in **chronological order** if multiple movements occur (e.g., \"The bird first takes flight, then soars in a circle, and finally lands on a branch\"). Since this represents the first and last frames of a video, refer to it as 'The video' instead of mentioning 'image' or 'frame.' Describe the entire video by combining details from both frames rather than separately. Focus on the setting and scenery rather than detailed subject descriptions. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B.' Return only the description.\n\nIf the subject in the video has no movement, please briefly mention that without going into too much detail.\n\nPlease only describe the content of the video. Don't mention the details of the subject's appearance unless you need to differentiate between multiple subjects by their appearance. Clearly describe the subject's motion.\n\nAvoid abstract descriptions, such as \"The car maintains a low, sleek profile as it maneuvers the bend, emphasizing its speed and agility\" and \"emphasizing its speed and agility as it maneuvers through the turn.\"\n\nBelow are detailed instructions:\n\nDescribe **individual subject actions** with clarity, specifying how they move rather than using generic descriptions. For example, instead of “a person is running,” say “a runner sprints across the finish line.”\n\nIf the subject interacts with an **object**, specify the type of interaction and its effect. Instead of “a person is working,” say “a construction worker operates a jackhammer, breaking up the pavement.”\n\nIf there are **interactions between subjects**, describe the nature of their relationship and movements relative to each other. Instead of “people are fighting,” say “two boxers exchange blows in the ring, circling each other cautiously.”\n\nIf there are collective behaviors for a group of subjects, describe **group activities** with specificity. Instead of “birds are flying,” say “a flock of geese flies in a V-formation across the horizon.” Instead of “people are walking,” say “a crowd of protesters marches down the street, carrying signs and banners.” Clearly convey the type of group, their coordinated actions, and any notable patterns in their movement.\n\nPlease note that the video features salient **human** subjects, so the description should focus on their motion and dynamics.\n\nIn addition, the human-written caption below already describes the subjects (if any) in this video but does not comprehensively capture their motion. Use this caption as a reference or starting point to draft the description, but do not rely on it completely. Expand or refine it to fully capture the subjects' motion and dynamics. However, if the caption does not mention any subjects, do not add any description—simply note this briefly in the description. Please note that this caption is only to help you clarify which subject's motion needs to be analyzed, rather than adding more content based on this description.\n\nHuman-Written Caption: **The video features a woman standing prominently on a skateboard in an indoor setting. She is wearing a white hat and light-colored clothing, with short hair. The woman appears to be in a relaxed stance, facing forward. Behind her, there is a display of skateboards mounted on the wall, adding context to the environment. The lighting highlights the woman and the skateboards, suggesting a focus on skateboarding or a related activity. Additionally, a drone-like device appears in the air later in the video.**",
"timestamp": "2025-03-19T10:37:53.434251",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"status": "rejected"
},
"scene": {
"caption_data": {
"caption_rating_score": null,
"feedback_is_needed": false,
"feedback_rating_score": null,
"final_caption": "The video depicts an indoor setting resembling a modern, industrial-style loft with a skateboarding theme. The environment features a wooden ramp prominently placed in the center, surrounded by various skateboards mounted on the walls. The space is illuminated by natural light streaming through large windows, casting shadows and creating a warm ambiance. The interior includes eclectic elements such as neon signs and a mix of furniture, contributing to a creative and urban atmosphere. The time of day appears to be late afternoon, as indicated by the angle and quality of the light. The scene remains consistent throughout, maintaining a focus on the unique combination of skate culture and contemporary design.",
"final_feedback": null,
"gpt_caption": null,
"gpt_caption_llm": null,
"gpt_caption_prompt": null,
"gpt_feedback": null,
"gpt_feedback_llm": null,
"gpt_feedback_prompt": null,
"initial_caption_rating_score": 5,
"initial_feedback": null,
"original_annotator": null,
"pre_caption": "The video depicts an indoor setting resembling a modern, industrial-style loft with a skateboarding theme. The environment features a wooden ramp prominently placed in the center, surrounded by various skateboards mounted on the walls. The space is illuminated by natural light streaming through large windows, casting shadows and creating a warm ambiance. The interior includes eclectic elements such as neon signs and a mix of furniture, contributing to a creative and urban atmosphere. The time of day appears to be late afternoon, as indicated by the angle and quality of the light. The scene remains consistent throughout, maintaining a focus on the unique combination of skate culture and contemporary design.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the overall scene, including the point of view, environment, setting, time of day, and notable visual elements. If subjects are present, ensure their placement and context complement the scene without excessive detail. Since this represents the first and last frames of a video, refer to it as 'The video' instead of mentioning 'image' or 'frame.' Describe the entire video by combining details from both frames rather than separately. Focus on the setting and scenery rather than detailed subject descriptions. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B.' Return only the description.\n\nIf relevant, indicate the **point of view**, such as first-person, drone shot, or dashcam, and describe how it influences the viewer’s perception. Specify the **setting** by clearly identifying whether it is indoors or outdoors, using precise language. If the location is known, state it explicitly (e.g., \"Times Square\" or \"Tokyo subway station\"). Otherwise, describe defining features such as “a narrow alley with graffiti-covered walls” or “a vast desert with rolling dunes.” Mention the **time of day** and any notable **architectural or natural features**, such as buildings, roads, forests, or bodies of water. Include relevant **weather conditions** if applicable, like “a rainy street with wet pavement reflecting city lights” or “a snowy mountain pass covered in thick fog.” For indoor settings, describe key **furniture or props** that establish the environment, such as “a wooden desk cluttered with books and a vintage lamp.” If notable, mention the **style** of the scene, such as a monochromatic color scheme or a vibrant carnival with neon lights. \n\nIf the scene involves **motion or changes**, describe natural elements like wind blowing through trees or waves crashing against the shore, as well as human-made movements such as traffic flowing on a highway or pedestrians walking along a busy street. If **scene transitions** occur, explain how they happen, whether through changes in time (e.g., “The shot transitions from day to night”) or movement-based shifts (e.g., “The camera pans to reveal an alley behind the main street”). Use precise language to refer to different elements within the scene, prioritizing the most prominent details and organizing descriptions logically from the overall setting to more specific features.\n\nIn addition, you do not need to infer the camera's point of view, as this information is already provided. Please integrate the following point of view information into your caption:\n\nPoint of View Information: \"**An objective POV shot where the camera captures the scene from a detached, observational perspective. (no need to mention).**\"",
"review_timestamp": "2025-05-21T06:42:25.426343",
"reviewer": "Shihang Zhu",
"reviewer_double_check": true,
"timestamp": "2025-03-12T23:18:35.784935",
"user": "Yuhan Huang",
"workflow_type": "perfect_precaption"
},
"negative_example": null,
"status": "approved"
},
"spatial": {
"caption_data": {
"caption_rating_score": 4,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "This video presents a wide shot, maintaining this framing throughout with a slight, slow push-in towards the subject. Initially, a woman is positioned in the center of the frame, standing in the middle ground atop a wooden ramp. This ramp is in the bottom middle ground. Initially, a grated floor is visible in the bottom middle ground but then exits the frame. An out-of-focus green object dominates the right foreground initially, but later exits the frame. In the background, behind the woman and slightly to her left, several skateboards are mounted on the wall. Further left in the background, a neon sign is visible. The surrounding space in the middle ground and background contains various furniture and objects. A grated ceiling structure is visible at the top. The camera remains at the same height as the woman. As the shot progresses, the camera moves slightly forward. Towards the end, a small drone flies into the frame from the top left, positioning itself close to the woman in the middle ground.",
"final_feedback": "The ramp is positioned in the middle ground. Initially, a grated floor is visible in the bottom middle ground but then exits the frame. The camera is not at the woman's eye level; it is at the same height as the subject. Additionally, the small drone is close to the woman in the middle ground, not in the background.",
"gpt_caption": "This video presents a wide shot, maintaining this framing throughout with a slight, slow push-in towards the subject. Initially, a woman is positioned in the center of the frame, standing in the middle ground atop a wooden ramp. This ramp originates from the bottom center foreground and extends into the middle ground. Initially, a grated floor is visible in the bottom middle ground but then exits the frame. An out-of-focus green object dominates the right foreground. In the background, behind the woman and slightly to her left, several skateboards are mounted on the wall. Further left in the background, a neon sign is visible. The surrounding space in the middle ground and background contains various furniture and objects. A grated ceiling structure is visible at the top. The camera remains at the same height as the woman. As the shot progresses, the camera moves slightly forward. Towards the end, a small drone flies into the frame from the top left background, positioning itself close to the woman in the middle ground.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The ramp is positioned in the middle ground. Initially, a grated floor is visible in the bottom middle ground but then exits the frame. The camera is not at the woman's eye level; it is at the same height as the subject. Additionally, the small drone is close to the woman in the middle ground, not in the background.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "The ramp is in the middle ground.A grated floor is in the bottom middle ground initially, then exits the frame. not at the woman's eye level, just at the same height as the subject. a small drone is close to the woman in the middle ground, not background.",
"original_annotator": null,
"pre_caption": "This video presents a wide shot, maintaining this framing throughout with a slight, slow push-in towards the subject. Initially, a woman is positioned in the center of the frame, standing in the middle ground atop a wooden ramp. This ramp originates from the bottom center foreground and extends into the middle ground. An out-of-focus green object dominates the right foreground. In the background, behind the woman and slightly to her left, several skateboards are mounted on the wall. Further left in the background, a neon sign is visible. The surrounding space in the middle ground and background contains various furniture and objects. A grated floor occupies the foreground, while a grated ceiling structure is visible at the top. The camera remains at the woman's eye level. As the shot progresses, the camera moves slightly forward. Towards the end, a small drone flies into the frame from the top left background, positioning itself above the woman in the background.",
"pre_caption_llm": "gemini-2.5-pro-preview-03-25",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Analyze the subjects and elements in this video and provide a concise yet informative description of how subjects and elements are framed within the scene, including their **shot size, position, and changes** within the frame. Your goal is to describe the **spatial framing and dynamics** within the shot. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B'. Return only the description.\n\nFirst, specify the **shot size** based on the subject's size in the frame if major subjects are present. If the shot size is unclear, describe how much of the subject is visible. If no major subject exists (e.g., a scenery shot), describe the shot size in relation to the scenery.\n\nNext, describe the **spatial position of subjects and elements in the video**, if relevant. Indicate their approximate **2D position** within the frame using terms like **left, right, bottom left, bottom right, top right, top left, bottom, top, or center**. Additionally, describe their **3D position** within the scene as **foreground, middle ground, or background**. Analyze as many elements as possible, and for each element mentioned, provide both its **2D and 3D position**.\n\nFinally, describe the **camera’s height relative to the subject**, if relevant. Indicate whether the camera is positioned at the subject's height, above them, or below them. We already have this information provided at the end. If it’s not provided, try to describe it by yourself.\n\nIf **shot size or spatial position** changes, describe how these transitions occur clearly, specifying both the **initial and final states**.\n\nIn addition, the human-written captions below already describe the subjects (if any) and the scenery in this video but do not capture their spatial composition and movements. Use these captions as a reference, but do not rely on them completely. Your goal is to fully capture the spatial framing and movements in this video. Don’t write too much about the subject's or scenery's appearance.\n\nHuman-Written Description for Subjects: **The video features a woman standing confidently on a wooden ramp in an indoor setting. She has short, light-colored hair and is wearing a light-colored knit hat, tank top, and pants, suggesting a casual style. The woman appears focused and poised, with a relaxed posture. Surrounding her are various skateboards mounted on the wall, adding a dynamic and sporty atmosphere to the scene. The lighting casts distinct shadows, enhancing the visual depth and highlighting her presence in the space. At the end of the video, a drone enters the frame from the left side.**\nHuman-Written Description for Scenery: **The video depicts an indoor setting resembling a modern, industrial-style loft with a skateboarding theme. The environment features a wooden ramp prominently placed in the center, surrounded by various skateboards mounted on the walls. The space is illuminated by natural light streaming through large windows, casting shadows and creating a warm ambiance. The interior includes eclectic elements such as neon signs and a mix of furniture, contributing to a creative and urban atmosphere. The time of day appears to be late afternoon, as indicated by the angle and quality of the light. The scene remains consistent throughout, maintaining a focus on the unique combination of skate culture and contemporary design.**\n\nPlease note that the video features **salient human subjects**, so you should focus on describing the spatial framing and movements of them.\n\nShot Size Information: The video shows a wide shot of the subjects.\n\nCamera Height Relative to Subjects: The camera is positioned at the subject's level.",
"review_timestamp": "2025-05-29T02:02:49.111542",
"reviewer": "Shihang Zhu",
"reviewer_double_check": true,
"timestamp": "2025-04-15T10:15:55.850041",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"negative_example": null,
"status": "approved"
},
"subject": {
"caption_data": {
"caption_rating_score": 5,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The video features a woman standing confidently on a wooden ramp in an indoor setting. She has short, light-colored hair and is wearing a light-colored knit hat, tank top, and pants, suggesting a casual style. The woman appears focused and poised, with a relaxed posture. Surrounding her are various skateboards mounted on the wall, adding a dynamic and sporty atmosphere to the scene. The lighting casts distinct shadows, enhancing the visual depth and highlighting her presence in the space. At the end of the video, a drone enters the frame from the left side.",
"final_feedback": "The individual in the video is a woman wearing a light-colored knit hat. At the end of the video, a drone enters the frame from the left side.",
"gpt_caption": "The video features a woman standing confidently on a wooden ramp in an indoor setting. She has short, light-colored hair and is wearing a light-colored knit hat, tank top, and pants, suggesting a casual style. The woman appears focused and poised, with a relaxed posture. Surrounding her are various skateboards mounted on the wall, adding a dynamic and sporty atmosphere to the scene. The lighting casts distinct shadows, enhancing the visual depth and highlighting her presence in the space. At the end of the video, a drone enters the frame from the left side.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The individual in the video is a woman wearing a light-colored knit hat. At the end of the video, a drone enters the frame from the left side.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "这个人是女人，戴着浅色针织帽。视频最后有一个无人机从画面左边飞进画面",
"original_annotator": null,
"pre_caption": "The video features a person standing confidently on a wooden ramp in an indoor setting. This individual has short, light-colored hair and is wearing a light-colored tank top and pants, suggesting a casual style. The person appears to be focused and poised, with a relaxed posture. Surrounding them are various skateboards mounted on the wall, adding a dynamic and sporty atmosphere to the scene. The lighting casts distinct shadows, enhancing the visual depth and highlighting the person's presence in the space.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the subjects in this video. Keep the description concise and clear, focusing on subject types and visual attributes. Since this is the first and last frame of a video, refer to it as \"The video\" instead of mentioning \"image\" or \"frame.\" Describe the entire video by combining details from both frames, rather than describing them separately. Don't mention the background or motion unless it's necessary to distinguish subjects by location, action, or relationships. You must avoid describing what is not visible or what you are unsure about. You must use simple, natural English and ensure the description is a clear, concise, and coherent paragraph that highlights the most essential details. You must avoid subjective adjectives that convey emotions. Whenever you mention a subject, please describe its key visual attributes in detail. Return only the description.\n\nClearly identify each subject’s type, using precise terms such as “man,” “woman,” “dog,” “car,” or “tree,” rather than vague words like “thing” or “item.” If the subject type is ambiguous, use your best judgment and briefly explain your reasoning.\n\nDescribe key visual attributes with specific and descriptive language. For people, include details such as clothing color and style, skin tone, hairstyle, facial hair, age (if discernible), gender, ethnicity (if relevant and clear), and facial expression. For objects, describe their color, material, shape, and distinguishing features like texture or markings. Additionally, note the subject’s pose and orientation within the frame, such as standing, sitting, walking, or facing a certain direction. Pay attention to any objects that are not in their usual state, like a tilted lamp or an open book lying face down.\n\nIf there are multiple subjects to describe, ensure clarity in referring to each. The simplest way is by type, such as “the man,” “the dog,” or “the tree.” If multiple subjects belong to the same category, distinguish them using unique appearance traits (e.g., “the woman in the red dress,” “the man with the beard”), location within the scene (e.g., “the man on the left,” “the car in the midground”), actions (e.g., “the child playing with a ball,” “the bird flying”), or relationships to each other (e.g., “the man next to the woman,” “the first man that enters the frame”). Also, when describing multiple subjects, the order in which they are mentioned matters. Prioritize based on relevance, starting with the largest or most centered subject. If the scene unfolds over time, describe subjects in the order they appear. If temporal order isn’t relevant, begin with the most visually striking or important subject before moving to less prominent ones. The goal is to provide enough detail so that anyone reading the description can easily identify each subject.\n\nPlease note that the video features salient **human** subjects, so the description should focus on them.",
"review_timestamp": "2025-05-05T02:58:46.468264",
"reviewer": "Zhiqiu Lin",
"reviewer_double_check": true,
"timestamp": "2025-03-20T09:57:18.035044",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"negative_example": null,
"status": "approved"
}
}
|
-2uIa-XMJC0.4.2.mp4
|
https://huggingface.co/datasets/zhiqiulin/video_captioning/resolve/main/-2uIa-XMJC0.4.2.mp4
|
{
"camera": {
"caption_data": {
"caption_rating_score": 5,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The video features an aerial-level camera that starts at a level angle and transitions to a low angle. It maintains deep focus with a large depth of field. The camera steadily moves forward, left, and downward, slightly tilting up and panning right to keep a skateboarder in frame, with some slight shaking.",
"final_feedback": "The camera steadily moves forward, left, and downward.",
"gpt_caption": "The video features an aerial-level camera that starts at a level angle and transitions to a low angle. It maintains deep focus with a large depth of field. The camera steadily moves forward, left, and downward, slightly tilting up and panning right to keep a skateboarder in frame, with some slight shaking.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The camera steadily moves forward, left, and downward.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 4,
"initial_feedback": "The camera steadily moves forward and downward这里修改为The camera steadily moves forward, left and downward",
"original_annotator": "Yuhan Huang",
"pre_caption": "The video features an aerial-level camera that starts at a level angle and transitions to a low angle. It maintains deep focus with a large depth of field. The camera steadily moves forward and downward, slightly tilting up and panning right to keep a skateboarder in frame, with some slight shaking.",
"pre_caption_llm": "gemini-2.5-pro-preview-05-06",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the **video’s and camera’s configuration**, covering **video speed, lens distortion, camera angle, camera height, movements (translation, rotation, steadiness, intensity, complexity), and focus (depth, focus plane, focus changes).** \n\nIf **video speed** is altered, specify the type, such as *time-lapse* (“Clouds move rapidly across the sky”), *slow-motion*, or *speed ramp* (a mix of fast and slow speeds). If the video is *time-reversed* or *stop-motion*, note this as well. \n\nIf **lens distortion** is present, describe the type and degree. For example, *fisheye distortion* creates extreme curvature, while *barrel distortion* causes mild outward bowing of straight lines near the edges. \n\nDescribe the **camera height** in relation to the ground, such as *eye-level, ground-level, overhead-level, or underwater*. If height changes due to movement, mention how it transitions. Similarly, specify the **camera angle**, such as *bird’s eye, high angle, low angle, or worm’s eye*, noting any shifts within the video. If a **Dutch angle** (tilted horizon) is present, indicate whether it remains fixed or varies due to camera rolling. \n\nIf discernible, describe the **camera focus and depth of field**. For example, *deep focus* keeps all elements sharp, while *shallow depth of field* blurs the background or foreground. If focus changes dynamically, note whether it’s a *rack focus* (shifting focus between subjects) or *focus tracking* (following a subject’s depth movement). If the video lacks realistic depth of field, describe whether it appears artificial, overly blurry, or lacking in cinematic focus effects. \n\nIf the **camera is static**, no further description is needed. If it moves, describe its **type, direction, and speed**. Specify movements such as *tracking (following a subject), arcing (circling around), dollying (moving forward/backward), trucking (moving left/right), panning (rotating horizontally), tilting (angling vertically), pedestaling (lifting/lowering), or rolling (rotating around the lens axis).* If motion intensity varies, distinguish between slow and rapid movements and present them in **temporal order** (e.g., *“The camera first pans right, then tilts upward to follow the subject”*). If motion appears **fragmented or random**, summarize it concisely rather than over-describing each small movement (e.g., *“The first-person camera moves erratically, frequently changing direction without a clear pattern”*).\n\nCrucially, instead of inferring these attributes from the video, we have already provided human-labeled ground truth for some of the elements specified above. You should directly use this information in your description and should not infer any details that are not already provided. Your description should be brief, and if anything is normal or unremarkable, you do not need to include it (e.g., if the video is at regular playback speed, there is no need to mention it).\n\n**Playback Speed:** The playback is forward at real-time speed without noticeable speed-up or slowdown. (no need to mention).\n\n**Lens Distortion:** No lens distortion (no need to mention).\n\n**Camera Height:** The camera is at an aerial level.\n\n**Camera Angle:** The camera angle is initially at a level angle (looking straight ahead) and then changes to a low angle (looking up from below) due to camera motion.\n\n**Camera Focus:** The camera uses a deep focus with a large depth of field.\n\n**Camera Motion:** The camera focuses on a skateboarder, steadily moving forward and downward while slightly tilting up and panning right to keep them in frame.\n\n**Camera Steadiness:** The camera movement is slightly unsteady with some shaking.\n\nIf possible, specify the subject that the camera focuses on when describing camera work. For instance, use 'focus on the man in the foreground' rather than 'focus on the foreground.' Likewise, if the camera follows a subject, avoid the generic phrase 'tracking the subject(s).' Instead, identify the subject and describe the specific type of tracking.",
"review_timestamp": "2025-05-29T02:06:02.132528",
"reviewer": "Shihang Zhu",
"reviewer_double_check": false,
"timestamp": "2025-05-29T02:07:37.813650",
"user": "Shihang Zhu",
"workflow_type": "improved_precaption"
},
"negative_example": {
"caption_rating_score": null,
"feedback_is_needed": false,
"feedback_rating_score": null,
"final_caption": "The video features an aerial-level camera that starts at a level angle and transitions to a low angle. It maintains deep focus with a large depth of field. The camera steadily moves forward and downward, slightly tilting up and panning right to keep a skateboarder in frame, with some slight shaking.",
"final_feedback": null,
"gpt_caption": null,
"gpt_caption_llm": null,
"gpt_caption_prompt": null,
"gpt_feedback": null,
"gpt_feedback_llm": null,
"gpt_feedback_prompt": null,
"initial_caption_rating_score": 5,
"initial_feedback": null,
"pre_caption": "The video features an aerial-level camera that starts at a level angle and transitions to a low angle. It maintains deep focus with a large depth of field. The camera steadily moves forward and downward, slightly tilting up and panning right to keep a skateboarder in frame, with some slight shaking.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the **video’s and camera’s configuration**, covering **video speed, lens distortion, camera angle, camera height, movements (translation, rotation, steadiness, intensity, complexity), and focus (depth, focus plane, focus changes).** \n\nIf **video speed** is altered, specify the type, such as *time-lapse* (“Clouds move rapidly across the sky”), *slow-motion*, or *speed ramp* (a mix of fast and slow speeds). If the video is *time-reversed* or *stop-motion*, note this as well. \n\nIf **lens distortion** is present, describe the type and degree. For example, *fisheye distortion* creates extreme curvature, while *barrel distortion* causes mild outward bowing of straight lines near the edges. \n\nDescribe the **camera height** in relation to the ground, such as *eye-level, ground-level, overhead-level, or underwater*. If height changes due to movement, mention how it transitions. Similarly, specify the **camera angle**, such as *bird’s eye, high angle, low angle, or worm’s eye*, noting any shifts within the video. If a **Dutch angle** (tilted horizon) is present, indicate whether it remains fixed or varies due to camera rolling. \n\nIf discernible, describe the **camera focus and depth of field**. For example, *deep focus* keeps all elements sharp, while *shallow depth of field* blurs the background or foreground. If focus changes dynamically, note whether it’s a *rack focus* (shifting focus between subjects) or *focus tracking* (following a subject’s depth movement). If the video lacks realistic depth of field, describe whether it appears artificial, overly blurry, or lacking in cinematic focus effects. \n\nIf the **camera is static**, no further description is needed. If it moves, describe its **type, direction, and speed**. Specify movements such as *tracking (following a subject), arcing (circling around), dollying (moving forward/backward), trucking (moving left/right), panning (rotating horizontally), tilting (angling vertically), pedestaling (lifting/lowering), or rolling (rotating around the lens axis).* If motion intensity varies, distinguish between slow and rapid movements and present them in **temporal order** (e.g., *“The camera first pans right, then tilts upward to follow the subject”*). If motion appears **fragmented or random**, summarize it concisely rather than over-describing each small movement (e.g., *“The first-person camera moves erratically, frequently changing direction without a clear pattern”*).\n\nCrucially, instead of inferring these attributes from the video, we have already provided human-labeled ground truth for some of the elements specified above. You should directly use this information in your description and should not infer any details that are not already provided. Your description should be brief, and if anything is normal or unremarkable, you do not need to include it (e.g., if the video is at regular playback speed, there is no need to mention it).\n\n**Playback Speed:** A regular-speed video where events play in real-time without acceleration or slowing. (no need to mention).\n\n**Lens Distortion:** No lens distortion (no need to mention).\n\n**Camera Height:** The camera is at an aerial level.\n\n**Camera Angle:** The camera angle is initially at a level angle (looking straight ahead) and then changes to a low angle (looking up from below) due to camera motion.\n\n**Camera Focus:** The camera uses a deep focus with a large depth of field.\n\n**Camera Motion:** The camera focuses on a skateboarder, steadily moving forward and downward while slightly tilting up and panning right to keep them in frame.\n\n**Camera Steadiness:** The camera movement is slightly unsteady with some shaking.",
"timestamp": "2025-03-21T10:27:12.785258",
"user": "Yuhan Huang",
"workflow_type": "perfect_precaption"
},
"status": "rejected"
},
"motion": {
"caption_data": {
"caption_rating_score": 3,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "A woman is standing on a skateboard, positioned on the ledge of a tall building. She is wearing a white tank top, light-colored pants, and a knitted cap, highlighting her focus on safety and style. Notably, she has a prosthetic right leg. She is balanced on the skateboard, which is oriented parallel to the building's windows, and is making a slight forward movement. The background shows a cityscape with various buildings and a clear sky, indicating a high altitude. The woman faces forward, with a confident posture.",
"final_feedback": "The woman on the skateboard is making a slight forward movement.",
"gpt_caption": "A woman is standing on a skateboard, positioned on the ledge of a tall building. She is wearing a white tank top, light-colored pants, and a knitted cap, highlighting her focus on safety and style. Notably, she has a prosthetic right leg. She is balanced on the skateboard, which is oriented parallel to the building's windows, and is making a slight forward movement. The background shows a cityscape with various buildings and a clear sky, indicating a high altitude. The woman maintains a steady stance, facing forward, with a confident posture.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The woman on the skateboard is making a slight forward movement.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 4,
"initial_feedback": "滑板上的女子有轻微往前移动",
"original_annotator": null,
"pre_caption": "A woman is standing on a skateboard, positioned on the ledge of a tall building. She is wearing a white tank top, light-colored pants, and a knitted cap, highlighting her focus on safety and style. Notably, she has a prosthetic right leg. She is balanced on the skateboard, which is oriented parallel to the building's windows. The background shows a cityscape with various buildings and a clear sky, indicating a high altitude. The woman maintains a steady stance, facing forward, with a confident posture.",
"pre_caption_llm": "tarsier-recap-7b",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the subject’s motion in this video, ensuring actions are presented in **chronological order** if multiple movements occur (e.g., \"The bird first takes flight, then soars in a circle, and finally lands on a branch\"). Since this represents the first and last frames of a video, refer to it as 'The video' instead of mentioning 'image' or 'frame.' Describe the entire video by combining details from both frames rather than separately. Focus on the setting and scenery rather than detailed subject descriptions. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B.' Return only the description.\n\nIf the subject in the video has no movement, please briefly mention that without going into too much detail.\n\nPlease only describe the content of the video. Don't mention the details of the subject's appearance unless you need to differentiate between multiple subjects by their appearance. Clearly describe the subject's motion.\n\nAvoid abstract descriptions, such as \"The car maintains a low, sleek profile as it maneuvers the bend, emphasizing its speed and agility\" and \"emphasizing its speed and agility as it maneuvers through the turn.\"\n\nBelow are detailed instructions:\n\nDescribe **individual subject actions** with clarity, specifying how they move rather than using generic descriptions. For example, instead of “a person is running,” say “a runner sprints across the finish line.”\n\nIf the subject interacts with an **object**, specify the type of interaction and its effect. Instead of “a person is working,” say “a construction worker operates a jackhammer, breaking up the pavement.”\n\nIf there are **interactions between subjects**, describe the nature of their relationship and movements relative to each other. Instead of “people are fighting,” say “two boxers exchange blows in the ring, circling each other cautiously.”\n\nIf there are collective behaviors for a group of subjects, describe **group activities** with specificity. Instead of “birds are flying,” say “a flock of geese flies in a V-formation across the horizon.” Instead of “people are walking,” say “a crowd of protesters marches down the street, carrying signs and banners.” Clearly convey the type of group, their coordinated actions, and any notable patterns in their movement.\n\nPlease note that the video features salient **human** subjects, so the description should focus on their motion and dynamics.\n\nIn addition, the human-written caption below already describes the subjects (if any) in this video but does not comprehensively capture their motion. Use this caption as a reference or starting point to draft the description, but do not rely on it completely. Expand or refine it to fully capture the subjects' motion and dynamics. However, if the caption does not mention any subjects, do not add any description—simply note this briefly in the description. Please note that this caption is only to help you clarify which subject's motion needs to be analyzed, rather than adding more content based on this description.\n\nHuman-Written Caption: **The video features a woman standing on a skateboard, positioned on the ledge of a tall building. She is wearing a white tank top, light-colored pants, and a knitted cap, highlighting her focus on safety and style. Notably, she has a prosthetic right leg. She is balanced on the skateboard, which is oriented parallel to the building's windows. The setting is urban, with a cityscape visible in the reflection of the windows, indicating a high altitude. The woman maintains a steady stance, facing forward, with a confident posture.**",
"review_timestamp": "2025-05-27T03:40:47.837797",
"reviewer": "Shihang Zhu",
"reviewer_double_check": true,
"timestamp": "2025-03-19T10:42:04.630749",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"negative_example": null,
"status": "approved"
},
"scene": {
"caption_data": {
"caption_rating_score": 4,
"feedback_is_needed": true,
"feedback_rating_score": 3,
"final_caption": "The video, captured from a drone POV, showcases a futuristic urban landscape during daylight. The setting is outdoors, featuring a cityscape with modern skyscrapers and reflective glass facades. The entire cityscape is beautifully mirrored in the glass facades. The architecture is sleek and contemporary, with tall buildings dominating the skyline. The weather is partly cloudy. ",
"final_feedback": "The video, captured from a drone POV, showcases a futuristic urban landscape during daylight. The setting is outdoors, featuring a cityscape with modern skyscrapers and reflective glass facades. The architecture is sleek and contemporary, with tall buildings dominating the skyline. The entire cityscape is beautifully reflected through the glass facades of the buildings, creating an illusion of a water body within the city. The weather appears to be cloudy.",
"gpt_caption": "The video, captured from a drone POV, showcases a futuristic urban landscape during daylight. The setting is outdoors, featuring a cityscape with modern skyscrapers and reflective glass facades. The architecture is sleek and contemporary, with tall buildings dominating the skyline. The entire cityscape is beautifully mirrored in the glass facades, creating an illusion of a water body within the city. The weather is partly cloudy, adding depth and texture to the expansive view of the city below. The drone's perspective provides a sweeping view of the urban environment, emphasizing the verticality and modernity of the setting.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The entire cityscape is beautifully reflected through the glass facades of the buildings, creating an illusion of a water body within the city. The weather appears to be cloudy, adding a dynamic element to the scene.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "整个城市景观是通过反光玻璃建筑物的玻璃反射呈现的，反应城市中似乎有一片水域，天气多云",
"original_annotator": null,
"pre_caption": "The video, captured from a drone POV, showcases a futuristic urban landscape during daylight. The setting is outdoors, featuring a cityscape with modern skyscrapers and reflective glass facades. The architecture is sleek and contemporary, with tall buildings dominating the skyline. The scene includes a person positioned on a ledge of a high-rise building, adding a sense of scale and height to the environment. The sky is clear with a few clouds, enhancing the bright and expansive view of the city below. The drone's perspective provides a sweeping view of the urban environment, emphasizing the verticality and modernity of the setting.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the overall scene, including the point of view, environment, setting, time of day, and notable visual elements. If subjects are present, ensure their placement and context complement the scene without excessive detail. Since this represents the first and last frames of a video, refer to it as 'The video' instead of mentioning 'image' or 'frame.' Describe the entire video by combining details from both frames rather than separately. Focus on the setting and scenery rather than detailed subject descriptions. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B.' Return only the description.\n\nIf relevant, indicate the **point of view**, such as first-person, drone shot, or dashcam, and describe how it influences the viewer’s perception. Specify the **setting** by clearly identifying whether it is indoors or outdoors, using precise language. If the location is known, state it explicitly (e.g., \"Times Square\" or \"Tokyo subway station\"). Otherwise, describe defining features such as “a narrow alley with graffiti-covered walls” or “a vast desert with rolling dunes.” Mention the **time of day** and any notable **architectural or natural features**, such as buildings, roads, forests, or bodies of water. Include relevant **weather conditions** if applicable, like “a rainy street with wet pavement reflecting city lights” or “a snowy mountain pass covered in thick fog.” For indoor settings, describe key **furniture or props** that establish the environment, such as “a wooden desk cluttered with books and a vintage lamp.” If notable, mention the **style** of the scene, such as a monochromatic color scheme or a vibrant carnival with neon lights. \n\nIf the scene involves **motion or changes**, describe natural elements like wind blowing through trees or waves crashing against the shore, as well as human-made movements such as traffic flowing on a highway or pedestrians walking along a busy street. If **scene transitions** occur, explain how they happen, whether through changes in time (e.g., “The shot transitions from day to night”) or movement-based shifts (e.g., “The camera pans to reveal an alley behind the main street”). Use precise language to refer to different elements within the scene, prioritizing the most prominent details and organizing descriptions logically from the overall setting to more specific features.\n\nIn addition, you do not need to infer the camera's point of view, as this information is already provided. Please integrate the following point of view information into your caption:\n\nPoint of View Information: \"**A drone POV shot.**\"",
"review_timestamp": "2025-05-21T06:51:39.294711",
"reviewer": "Shihang Zhu",
"reviewer_double_check": true,
"timestamp": "2025-03-14T02:23:47.634738",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"negative_example": null,
"status": "approved"
},
"spatial": {
"caption_data": {
"caption_rating_score": null,
"feedback_is_needed": false,
"feedback_rating_score": null,
"final_caption": "The video presents a wide shot of a woman standing on a skateboard on the ledge of a tall building. Initially, the camera is positioned at her level, capturing her from the side as she stands centered in the frame, with the building's windows forming the backdrop. The windows reflect the cityscape, situating the scene in the foreground while the urban landscape occupies the background. As the video progresses, the camera shifts to a lower angle, now positioned below the subject. This change in perspective emphasizes her elevated position on the ledge. Throughout the video, the woman remains in the center of the frame, maintaining a steady stance on the skateboard. The camera movement from her level to below her enhances the sense of height and scale, underscoring her position against the expansive cityscape reflected in the windows behind her.",
"final_feedback": null,
"gpt_caption": null,
"gpt_caption_llm": null,
"gpt_caption_prompt": null,
"gpt_feedback": null,
"gpt_feedback_llm": null,
"gpt_feedback_prompt": null,
"initial_caption_rating_score": 5,
"initial_feedback": null,
"original_annotator": null,
"pre_caption": "The video presents a wide shot of a woman standing on a skateboard on the ledge of a tall building. Initially, the camera is positioned at her level, capturing her from the side as she stands centered in the frame, with the building's windows forming the backdrop. The windows reflect the cityscape, situating the scene in the foreground while the urban landscape occupies the background. As the video progresses, the camera shifts to a lower angle, now positioned below the subject. This change in perspective emphasizes her elevated position on the ledge. Throughout the video, the woman remains in the center of the frame, maintaining a steady stance on the skateboard. The camera movement from her level to below her enhances the sense of height and scale, underscoring her position against the expansive cityscape reflected in the windows behind her.",
"pre_caption_llm": "qwen2.5-vl-72b",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Analyze the subjects and elements in this video and provide a concise yet informative description of how subjects and elements are framed within the scene, including their **shot size, position, and changes** within the frame. Your goal is to describe the **spatial framing and dynamics** within the shot. Since this represents the first and last frames of a video, refer to it as 'The video' instead of mentioning 'image' or 'frame.' Describe the entire video by combining details from both frames rather than separately. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B'. Return only the description.\n\nFirst, specify the **shot size** based on the subject's size in the frame if major subjects are present. If the shot size is unclear, describe how much of the subject is visible. If no major subject exists (e.g., a scenery shot), describe the shot size in relation to the scenery.\n\nNext, describe the **spatial position of subjects and elements in the video**, if relevant. Indicate their approximate **2D position** within the frame using terms like **left, right, bottom left, bottom right, top right, top left, bottom, top, or center**. Additionally, describe their **3D position** within the scene as **foreground, middle ground, or background**. Analyze as many elements as possible, and for each element mentioned, provide both its **2D and 3D position**.\n\nFinally, describe the **camera’s height relative to the subject**, if relevant. Indicate whether the camera is positioned at the subject's height, above them, or below them. We already have this information provided at the end. If it’s not provided, try to describe it by yourself.\n\nIf **shot size or spatial position** changes, describe how these transitions occur clearly, specifying both the **initial and final states**.\n\nIn addition, the human-written captions below already describe the subjects (if any) and the scenery in this video but do not capture their spatial composition and movements. Use these captions as a reference, but do not rely on them completely. Your goal is to fully capture the spatial framing and movements in this video. Don’t write too much about the subject's or scenery's appearance.\n\nHuman-Written Description for Subjects: **The video features a woman standing on a skateboard, positioned on the ledge of a tall building. She is wearing a white tank top, light-colored pants, and a knitted cap, highlighting her focus on safety and style. Notably, she has a prosthetic right leg. She is balanced on the skateboard, which is oriented parallel to the building's windows. The setting is urban, with a cityscape visible in the reflection of the windows, indicating a high altitude. The woman maintains a steady stance, facing forward, with a confident posture.**\nHuman-Written Description for Scenery: **The video, captured from a drone POV, showcases a futuristic urban landscape during daylight. The setting is outdoors, featuring a cityscape with modern skyscrapers and reflective glass facades. The entire cityscape is beautifully mirrored in the glass facades. The architecture is sleek and contemporary, with tall buildings dominating the skyline. The weather is partly cloudy. **\n\nPlease note that the video features **salient human subjects**, so you should focus on describing the spatial framing and movements of them.\n\nShot Size Information: The video shows a wide shot of the subjects.\n\nCamera Height Relative to Subjects: The camera is initially positioned at the subject's level. It then changes to below the subject.",
"review_timestamp": "2025-05-06T00:40:13.328061",
"reviewer": "Siyuan Cen",
"reviewer_double_check": true,
"timestamp": "2025-03-21T09:53:47.399363",
"user": "Yuhan Huang",
"workflow_type": "perfect_precaption"
},
"negative_example": null,
"status": "approved"
},
"subject": {
"caption_data": {
"caption_rating_score": 4,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The video features a woman with light skin and a prosthetic left leg standing on a skateboard. She is dressed entirely in white, wearing a tank top, pants, and a beanie. Her reflection is visible in the window pane directly behind her.",
"final_feedback": "The reflection of the woman is not in the window pane to her left, but rather directly behind her.",
"gpt_caption": "The video features a woman with light skin and a prosthetic left leg standing on a skateboard. She is dressed entirely in white, wearing a tank top, pants, and a beanie. Her reflection is visible directly behind her.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The reflection of the woman is not in the window pane to her left, but rather directly behind her.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 4,
"initial_feedback": "Her reflection is not in the window pane to her left but directly behind her.",
"original_annotator": "Yuhan Huang",
"pre_caption": "The video features a woman with light skin and a prosthetic left leg standing on a skateboard. She is dressed entirely in white, wearing a tank top, pants, and a beanie. Her reflection is visible in the window pane directly to her left.",
"pre_caption_llm": "gemini-2.5-pro-preview-03-25",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the subjects in this video. Keep the description concise and clear, focusing on subject types and visual attributes. Since this is the first and last frame of a video, refer to it as \"The video\" instead of mentioning \"image\" or \"frame.\" Describe the entire video by combining details from both frames, rather than describing them separately. Don't mention the background or motion unless it's necessary to distinguish subjects by location, action, or relationships. You must avoid describing what is not visible or what you are unsure about. You must use simple, natural English and ensure the description is a clear, concise, and coherent paragraph that highlights the most essential details. You must avoid subjective adjectives that convey emotions. Whenever you mention a subject, please describe its key visual attributes in detail. Return only the one-paragraph video description without Markdown formatting or introductory text.\n\nClearly identify each subject’s type, using precise terms such as “man,” “woman,” “dog,” “car,” or “tree,” rather than vague words like “thing” or “item.” If the subject type is ambiguous, use your best judgment and briefly explain your reasoning.\n\nDescribe key visual attributes with specific and descriptive language. For people, include details such as clothing color and style, skin tone, hairstyle, facial hair, age (if discernible), gender, ethnicity (if relevant and clear), and facial expression. For objects, describe their color, material, shape, and distinguishing features like texture or markings. Additionally, note the subject’s pose and orientation within the frame, such as standing, sitting, walking, or facing a certain direction. Pay attention to any objects that are not in their usual state, like a tilted lamp or an open book lying face down.\n\nIf there are multiple subjects to describe, ensure clarity in referring to each. The simplest way is by type, such as “the man,” “the dog,” or “the tree.” If multiple subjects belong to the same category, distinguish them using unique appearance traits (e.g., “the woman in the red dress,” “the man with the beard”), location within the scene (e.g., “the man on the left,” “the car in the midground”), actions (e.g., “the child playing with a ball,” “the bird flying”), or relationships to each other (e.g., “the man next to the woman,” “the first man that enters the frame”). Also, when describing multiple subjects, the order in which they are mentioned matters. Prioritize based on relevance, starting with the largest or most centered subject. If the scene unfolds over time, describe subjects in the order they appear. If temporal order isn’t relevant, begin with the most visually striking or important subject before moving to less prominent ones. The goal is to provide enough detail so that anyone reading the description can easily identify each subject.\n\nPlease note that the video features salient **human** subjects, so the description should focus on them.\n\nPlease avoid using phrases like \"the first frame\" or \"the last frame\" in your description. Instead, refer to the entire sequence simply as \"the video.\" Your description should integrate observations from all frames into a cohesive, temporally and logically consistent narrative, rather than describing frames in isolation. Whenever you mention a subject, be sure to include detailed descriptions of its key visual attributes. The final output should be a single, fluent paragraph describing the video, with no Markdown formatting or introductory text.",
"review_timestamp": "2025-05-05T16:50:01.284330",
"reviewer": "Siyuan Cen",
"reviewer_double_check": false,
"timestamp": "2025-05-05T16:51:58.920877",
"user": "Siyuan Cen",
"workflow_type": "improved_precaption"
},
"negative_example": {
"caption_rating_score": null,
"feedback_is_needed": false,
"feedback_rating_score": null,
"final_caption": "The video features a woman with light skin and a prosthetic left leg standing on a skateboard. She is dressed entirely in white, wearing a tank top, pants, and a beanie. Her reflection is visible in the window pane directly to her left.",
"final_feedback": null,
"gpt_caption": null,
"gpt_caption_llm": null,
"gpt_caption_prompt": null,
"gpt_feedback": null,
"gpt_feedback_llm": null,
"gpt_feedback_prompt": null,
"initial_caption_rating_score": 5,
"initial_feedback": null,
"pre_caption": "The video features a woman with light skin and a prosthetic left leg standing on a skateboard. She is dressed entirely in white, wearing a tank top, pants, and a beanie. Her reflection is visible in the window pane directly to her left.",
"pre_caption_llm": "gemini-2.5-pro-preview-03-25",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the subjects in this video. Keep the description concise and clear, focusing on subject types and visual attributes. Since this is the first and last frame of a video, refer to it as \"The video\" instead of mentioning \"image\" or \"frame.\" Describe the entire video by combining details from both frames, rather than describing them separately. Don't mention the background or motion unless it's necessary to distinguish subjects by location, action, or relationships. You must avoid describing what is not visible or what you are unsure about. You must use simple, natural English and ensure the description is a clear, concise, and coherent paragraph that highlights the most essential details. You must avoid subjective adjectives that convey emotions. Whenever you mention a subject, please describe its key visual attributes in detail. Return only the description.\n\nClearly identify each subject’s type, using precise terms such as “man,” “woman,” “dog,” “car,” or “tree,” rather than vague words like “thing” or “item.” If the subject type is ambiguous, use your best judgment and briefly explain your reasoning.\n\nDescribe key visual attributes with specific and descriptive language. For people, include details such as clothing color and style, skin tone, hairstyle, facial hair, age (if discernible), gender, ethnicity (if relevant and clear), and facial expression. For objects, describe their color, material, shape, and distinguishing features like texture or markings. Additionally, note the subject’s pose and orientation within the frame, such as standing, sitting, walking, or facing a certain direction. Pay attention to any objects that are not in their usual state, like a tilted lamp or an open book lying face down.\n\nIf there are multiple subjects to describe, ensure clarity in referring to each. The simplest way is by type, such as “the man,” “the dog,” or “the tree.” If multiple subjects belong to the same category, distinguish them using unique appearance traits (e.g., “the woman in the red dress,” “the man with the beard”), location within the scene (e.g., “the man on the left,” “the car in the midground”), actions (e.g., “the child playing with a ball,” “the bird flying”), or relationships to each other (e.g., “the man next to the woman,” “the first man that enters the frame”). Also, when describing multiple subjects, the order in which they are mentioned matters. Prioritize based on relevance, starting with the largest or most centered subject. If the scene unfolds over time, describe subjects in the order they appear. If temporal order isn’t relevant, begin with the most visually striking or important subject before moving to less prominent ones. The goal is to provide enough detail so that anyone reading the description can easily identify each subject.\n\nPlease note that the video features salient **human** subjects, so the description should focus on them.\n\nSince this is the first and last frame of a video, refer to it as \"The video\" instead of mentioning \"image\" or \"frame.\" Don't use terms like \"the first frame\", \"the last frame\". Describe the entire video by combining details from both frames, rather than describing them separately. Whenever you mention a subject, please describe its key visual attributes in detail.",
"timestamp": "2025-04-15T22:50:54.699121",
"user": "Yuhan Huang",
"workflow_type": "perfect_precaption"
},
"status": "rejected"
}
}
|
-2uIa-XMJC0.6.11.mp4
|
https://huggingface.co/datasets/zhiqiulin/video_captioning/resolve/main/-2uIa-XMJC0.6.11.mp4
|
{
"camera": {
"caption_data": {
"caption_rating_score": 5,
"feedback_is_needed": true,
"feedback_rating_score": 4,
"final_caption": "The video is shot from an eye-level perspective with a generally level camera angle, occasionally tilting slightly up and down, creating brief high-angle views. It features a shallow depth of field, emphasizing the screen in the foreground. The camera is unsteady, with noticeable shaking, and intermittently tilts downward as the person interacts with a USB connection on the screen.",
"final_feedback": "The video focuses on the screen in the foreground.",
"gpt_caption": "The video is shot from an eye-level perspective with a generally level camera angle, occasionally tilting slightly up and down, creating brief high-angle views. It features a shallow depth of field, emphasizing the screen in the foreground. The camera is unsteady, with noticeable shaking, and intermittently tilts downward as the person interacts with a USB connection on the screen.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The video effectively focuses on the screen in the foreground, highlighting its importance in the scene.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 4,
"initial_feedback": "focusing on a screen in the foreground",
"original_annotator": null,
"pre_caption": "The video is shot from an eye-level perspective with a generally level camera angle, occasionally tilting slightly up and down, creating brief high-angle views. It features a shallow depth of field, focusing on the foreground. The camera is unsteady, with noticeable shaking, and intermittently tilts downward as the person interacts with a USB connection on the screen.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the **video’s and camera’s configuration**, covering **video speed, lens distortion, camera angle, camera height, movements (translation, rotation, steadiness, intensity, complexity), and focus (depth, focus plane, focus changes).** \n\nIf **video speed** is altered, specify the type, such as *time-lapse* (“Clouds move rapidly across the sky”), *slow-motion*, or *speed ramp* (a mix of fast and slow speeds). If the video is *time-reversed* or *stop-motion*, note this as well. \n\nIf **lens distortion** is present, describe the type and degree. For example, *fisheye distortion* creates extreme curvature, while *barrel distortion* causes mild outward bowing of straight lines near the edges. \n\nDescribe the **camera height** in relation to the ground, such as *eye-level, ground-level, overhead-level, or underwater*. If height changes due to movement, mention how it transitions. Similarly, specify the **camera angle**, such as *bird’s eye, high angle, low angle, or worm’s eye*, noting any shifts within the video. If a **Dutch angle** (tilted horizon) is present, indicate whether it remains fixed or varies due to camera rolling. \n\nIf discernible, describe the **camera focus and depth of field**. For example, *deep focus* keeps all elements sharp, while *shallow depth of field* blurs the background or foreground. If focus changes dynamically, note whether it’s a *rack focus* (shifting focus between subjects) or *focus tracking* (following a subject’s depth movement). If the video lacks realistic depth of field, describe whether it appears artificial, overly blurry, or lacking in cinematic focus effects. \n\nIf the **camera is static**, no further description is needed. If it moves, describe its **type, direction, and speed**. Specify movements such as *tracking (following a subject), arcing (circling around), dollying (moving forward/backward), trucking (moving left/right), panning (rotating horizontally), tilting (angling vertically), pedestaling (lifting/lowering), or rolling (rotating around the lens axis).* If motion intensity varies, distinguish between slow and rapid movements and present them in **temporal order** (e.g., *“The camera first pans right, then tilts upward to follow the subject”*). If motion appears **fragmented or random**, summarize it concisely rather than over-describing each small movement (e.g., *“The first-person camera moves erratically, frequently changing direction without a clear pattern”*).\n\nCrucially, instead of inferring these attributes from the video, we have already provided human-labeled ground truth for some of the elements specified above. You should directly use this information in your description and should not infer any details that are not already provided. Your description should be brief, and if anything is normal or unremarkable, you do not need to include it (e.g., if the video is at regular playback speed, there is no need to mention it).\n\n**Playback Speed:** A regular-speed video where events play in real-time without acceleration or slowing. (no need to mention).\n\n**Lens Distortion:** No lens distortion (no need to mention).\n\n**Camera Height:** The camera is at an eye level (above the waist).\n\n**Camera Angle:** The video maintains a generally level angle, but there is slight tilting down and up, creating brief moments of a high angle.\n\n**Camera Focus:** The camera uses a shallow depth of field. The camera is focused on the foreground.\n\n**Camera Motion:** The unsteady camera, capturing a screen from a first-person perspective, intermittently tilts downward as the person looks down to plug or unplug the USB connected to the screen.\n\n**Camera Steadiness:** The camera movement is unsteady with noticeable shaking.\n\nIf possible, specify the subject that the camera focuses on when describing camera work. For instance, use 'focus on the man in the foreground' rather than 'focus on the foreground.' Likewise, if the camera follows a subject, avoid the generic phrase 'tracking the subject(s).' Instead, identify the subject and describe the specific type of tracking.",
"review_timestamp": "2025-05-29T02:09:32.546654",
"reviewer": "Shihang Zhu",
"reviewer_double_check": true,
"timestamp": "2025-03-30T10:24:18.763685",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"negative_example": null,
"status": "approved"
},
"motion": {
"caption_data": {
"caption_rating_score": 4,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "A man is seen on a circular screen, speaking and moving his head slightly. The background shows a blurred cityscape with buildings. The man continues to speak and move his head, and a hand appears from the bottom of the screen, removing an orange USB drive from the screen's port, reinserting it, and then removing a different pink USB drive. The screen changes due to the USB drive being swapped. The man on the screen continues to speak, and a smaller circular frame with a Black man in a red hoodie appears at the bottom right of the screen.",
"final_feedback": "The hand removed an orange USB drive from the screen, reinserted it, and then removed a different pink USB drive.",
"gpt_caption": "A man is seen on a circular screen, speaking and moving his head slightly. The background shows a blurred cityscape with buildings. The man continues to speak and move his head as a hand appears from the bottom of the screen, removing an orange USB drive from the screen's port, reinserting it, and then removing a different pink USB drive. The screen's lights change colors, and the man on the screen continues to speak. A smaller circular frame with a Black man in a red hoodie appears at the bottom right of the screen.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The hand removed an orange USB drive from the screen, reinserted it, and then removed a different pink USB drive.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "手先拔掉了显示屏上的一个橘色U盘，又插回去，然后又拔掉了另一个粉色U盘",
"original_annotator": null,
"pre_caption": "A man is seen on a circular screen, speaking and moving his head slightly. The background shows a blurred cityscape with buildings. The man continues to speak and move his head, and a hand appears from the bottom of the screen, holding a USB drive. The hand plugs the USB drive into the screen's port, and the screen's lights change colors. The man on the screen continues to speak, and a smaller circular frame with a Black man in a red hoodie appears at the bottom right of the screen.",
"pre_caption_llm": "tarsier-recap-7b",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the subject’s motion in this video, ensuring actions are presented in **chronological order** if multiple movements occur (e.g., \"The bird first takes flight, then soars in a circle, and finally lands on a branch\"). Since this represents the first and last frames of a video, refer to it as 'The video' instead of mentioning 'image' or 'frame.' Describe the entire video by combining details from both frames rather than separately. Focus on the setting and scenery rather than detailed subject descriptions. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B.' Return only the description.\n\nIf the subject in the video has no movement, please briefly mention that without going into too much detail.\n\nPlease only describe the content of the video. Don't mention the details of the subject's appearance unless you need to differentiate between multiple subjects by their appearance. Clearly describe the subject's motion.\n\nAvoid abstract descriptions, such as \"The car maintains a low, sleek profile as it maneuvers the bend, emphasizing its speed and agility\" and \"emphasizing its speed and agility as it maneuvers through the turn.\"\n\nBelow are detailed instructions:\n\nDescribe **individual subject actions** with clarity, specifying how they move rather than using generic descriptions. For example, instead of “a person is running,” say “a runner sprints across the finish line.”\n\nIf the subject interacts with an **object**, specify the type of interaction and its effect. Instead of “a person is working,” say “a construction worker operates a jackhammer, breaking up the pavement.”\n\nIf there are **interactions between subjects**, describe the nature of their relationship and movements relative to each other. Instead of “people are fighting,” say “two boxers exchange blows in the ring, circling each other cautiously.”\n\nIf there are collective behaviors for a group of subjects, describe **group activities** with specificity. Instead of “birds are flying,” say “a flock of geese flies in a V-formation across the horizon.” Instead of “people are walking,” say “a crowd of protesters marches down the street, carrying signs and banners.” Clearly convey the type of group, their coordinated actions, and any notable patterns in their movement.\n\nIn addition, the human-written caption below already describes the subjects (if any) in this video but does not comprehensively capture their motion. Use this caption as a reference or starting point to draft the description, but do not rely on it completely. Expand or refine it to fully capture the subjects' motion and dynamics. However, if the caption does not mention any subjects, do not add any description—simply note this briefly in the description. Please note that this caption is only to help you clarify which subject's motion needs to be analyzed, rather than adding more content based on this description.\n\nHuman-Written Caption: **The video shows a man with short, brown hair, wearing a gray hoodie and a necklace, appearing within a circular frame that resembles a digital interface. A smaller circular frame appears, featuring a Black man in a red hoodie, with a focused expression. The setting suggests a virtual or digital communication scenario, viewed from a first-person perspective as the red hoodie man's hand briefly appears at the bottom of the screen. **\n\nThis caption is only to help you learn the video thoroughly, so don't rely on it too much. Focus on the subject’s motion!! \n",
"review_timestamp": "2025-05-07T02:27:46.093440",
"reviewer": "Siyuan Cen",
"reviewer_double_check": true,
"timestamp": "2025-03-19T10:46:54.907719",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"negative_example": null,
"status": "approved"
},
"scene": {
"caption_data": {
"caption_rating_score": 5,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The video presents a first-person POV shot set indoors, featuring a modern, high-tech environment with floor-to-ceiling windows. A large, transparent screen displays a video call interface with a circular view of an outdoor urban setting, showing buildings and a clear sky. The time of day appears to be daytime, with bright natural light illuminating the scene. The screen also includes a smaller circular inset showing another person, suggesting a conversation. The room is sleek and futuristic, with colorful walls in shades of blue, pink, and white, and a low shelving unit on the left side, emphasizing digital communication technology.",
"final_feedback": "The background of the room features floor-to-ceiling windows, with a low shelving unit on the left side. The walls of the room are colorful, predominantly in shades of blue, pink, and white.",
"gpt_caption": "The video presents a first-person POV shot set indoors, featuring a modern, high-tech environment with floor-to-ceiling windows. A large, transparent screen displays a video call interface with a circular view of an outdoor urban setting, showing buildings and a clear sky. The time of day appears to be daytime, with bright natural light illuminating the scene. The screen also includes a smaller circular inset showing another person, suggesting a conversation. The room is sleek and futuristic, with colorful walls in shades of blue, pink, and white, and a low shelving unit on the left side, emphasizing digital communication technology.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The background of the room features floor-to-ceiling windows, with a low shelving unit on the left side. The walls of the room are colorful, predominantly in shades of blue, pink, and white.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "背景中显示所处房间有着落地窗，左边有矮的置物架，背景房间的墙壁是彩色的，主要有蓝色粉色白色",
"original_annotator": null,
"pre_caption": "The video presents a first-person POV shot set indoors, featuring a modern, high-tech environment. A large, transparent screen displays a video call interface with a circular view of an outdoor urban setting, showing buildings and a clear sky. The time of day appears to be daytime, with bright natural light illuminating the scene. The screen also includes a smaller circular inset showing another person, suggesting a conversation. The setting is sleek and futuristic, with a focus on digital communication technology.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the overall scene, including the point of view, environment, setting, time of day, and notable visual elements. If subjects are present, ensure their placement and context complement the scene without excessive detail. Since this represents the first and last frames of a video, refer to it as 'The video' instead of mentioning 'image' or 'frame.' Describe the entire video by combining details from both frames rather than separately. Focus on the setting and scenery rather than detailed subject descriptions. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B.' Return only the description.\n\nIf relevant, indicate the **point of view**, such as first-person, drone shot, or dashcam, and describe how it influences the viewer’s perception. Specify the **setting** by clearly identifying whether it is indoors or outdoors, using precise language. If the location is known, state it explicitly (e.g., \"Times Square\" or \"Tokyo subway station\"). Otherwise, describe defining features such as “a narrow alley with graffiti-covered walls” or “a vast desert with rolling dunes.” Mention the **time of day** and any notable **architectural or natural features**, such as buildings, roads, forests, or bodies of water. Include relevant **weather conditions** if applicable, like “a rainy street with wet pavement reflecting city lights” or “a snowy mountain pass covered in thick fog.” For indoor settings, describe key **furniture or props** that establish the environment, such as “a wooden desk cluttered with books and a vintage lamp.” If notable, mention the **style** of the scene, such as a monochromatic color scheme or a vibrant carnival with neon lights. \n\nIf the scene involves **motion or changes**, describe natural elements like wind blowing through trees or waves crashing against the shore, as well as human-made movements such as traffic flowing on a highway or pedestrians walking along a busy street. If **scene transitions** occur, explain how they happen, whether through changes in time (e.g., “The shot transitions from day to night”) or movement-based shifts (e.g., “The camera pans to reveal an alley behind the main street”). Use precise language to refer to different elements within the scene, prioritizing the most prominent details and organizing descriptions logically from the overall setting to more specific features.\n\nSince this is the first and last frame of a video, refer to it as \"The video\" instead of mentioning \"image\" or \"frame.\" Don't use terms like \"the first frame\", \"the last frame\". Describe the entire video by combining details from both frames, rather than describing them separately.\n\nIn addition, you do not need to infer the camera's point of view, as this information is already provided. Please integrate the following point of view information into your caption:\n\nPoint of View Information: \"**A first-person POV shot.**\"",
"review_timestamp": "2025-05-21T06:53:21.802805",
"reviewer": "Shihang Zhu",
"reviewer_double_check": true,
"timestamp": "2025-03-19T05:05:46.537457",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"negative_example": null,
"status": "approved"
},
"spatial": {
"caption_data": {
"caption_rating_score": 4,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The video captures a first-person perspective within a modern, high-tech indoor setting. In the center of the scene, a large, transparent screen dominates the foreground, displaying a circular digital interface with a man in a gray hoodie and necklace. This man occupies the central area of the circular frame, which is positioned in the middle of the screen. The background features a colorful wall with shades of blue, pink, and white, occupying a significant portion of the scene. On the left side of the screen, in the middle ground, a low shelving unit holds various items, including what appears to be a drone. Additionally, there is a partially visible floor-to-ceiling window on the left side of the background. As the video progresses, a hand wearing a red sleeve enters from the bottom of the frame, interacting with the screen's controls located at the bottom edge. Simultaneously, a smaller circular inset appears near the bottom right corner of the screen, showing another individual in a red hoodie. The camera remains at eye level throughout, maintaining a consistent viewpoint relative to the subjects and objects in the scene. The overall spatial arrangement emphasizes the interaction between the viewer and the digital interface, with the screen acting as the focal point.",
"final_feedback": "The colorful wall occupies a significant portion of the background, not just the right side. The low shelving unit is positioned in the middle ground on the left side and does not contain a vinyl record player; it appears to be a drone instead. Additionally, there is a partially visible floor-to-ceiling window on the left side of the background.",
"gpt_caption": "The video captures a first-person perspective within a modern, high-tech indoor setting. In the center of the scene, a large, transparent screen dominates the middle ground, displaying a circular digital interface with a man in a gray hoodie and necklace. This man occupies the central area of the circular frame, which is positioned in the middle of the screen. The background features a colorful wall with shades of blue, pink, and white, occupying a significant portion of the scene. On the left side of the screen, in the middle ground, a low shelving unit holds various items, including what appears to be a drone. Additionally, a partially visible floor-to-ceiling window is present on the left side of the background. As the video progresses, a hand wearing a red sleeve enters from the bottom of the frame, interacting with the screen's controls located at the bottom edge. Simultaneously, a smaller circular inset appears near the bottom right corner of the screen, showing another individual in a red hoodie. The camera remains at eye level throughout, maintaining a consistent viewpoint relative to the subjects and objects in the scene. The overall spatial arrangement emphasizes the interaction between the viewer and the digital interface, with the screen acting as the focal point.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The colorful wall occupies a significant portion of the background, not just the right side. The low shelving unit is positioned in the middle ground on the left side and does not contain a vinyl record player; it appears to be a drone instead. Additionally, there is a partially visible floor-to-ceiling window on the left side of the background.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 2,
"initial_feedback": "彩色的墙在背景中占了大部分，不是只在右侧。矮架在中景左侧，没有黑胶唱片，应该是个无人机。背景中左侧有部分可见的落地窗。",
"original_annotator": null,
"pre_caption": "The video captures a first-person perspective within a modern, high-tech indoor setting. In the center of the scene, a large, transparent screen dominates the middle ground, displaying a circular digital interface with a man in a gray hoodie and necklace. This man occupies the central area of the circular frame, which is positioned in the middle of the screen. To the right of the screen, in the background, there is a colorful wall with shades of blue, pink, and white. On the left side of the screen, in the foreground, a low shelving unit holds various items, including what appears to be a vinyl record player. As the video progresses, a hand wearing a red sleeve enters from the bottom of the frame, interacting with the screen's controls located at the bottom edge. Simultaneously, a smaller circular inset appears near the bottom right corner of the screen, showing another individual in a red hoodie. The camera remains at eye level throughout, maintaining a consistent viewpoint relative to the subjects and objects in the scene. The overall spatial arrangement emphasizes the interaction between the viewer and the digital interface, with the screen acting as the focal point.",
"pre_caption_llm": "qwen2.5-vl-72b",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Analyze the subjects and elements in this video and provide a concise yet informative description of how subjects and elements are framed within the scene, including their **shot size, position, and changes** within the frame. Your goal is to describe the **spatial framing and dynamics** within the shot. Since this represents the first and last frames of a video, refer to it as 'The video' instead of mentioning 'image' or 'frame.' Describe the entire video by combining details from both frames rather than separately. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B'. Return only the description.\n\nFirst, specify the **shot size** based on the subject's size in the frame if major subjects are present. If the shot size is unclear, describe how much of the subject is visible. If no major subject exists (e.g., a scenery shot), describe the shot size in relation to the scenery.\n\nNext, describe the **spatial position of subjects and elements in the video**, if relevant. Indicate their approximate **2D position** within the frame using terms like **left, right, bottom left, bottom right, top right, top left, bottom, top, or center**. Additionally, describe their **3D position** within the scene as **foreground, middle ground, or background**. Analyze as many elements as possible, and for each element mentioned, provide both its **2D and 3D position**.\n\nFinally, describe the **camera’s height relative to the subject**, if relevant. Indicate whether the camera is positioned at the subject's height, above them, or below them. We already have this information provided at the end. If it’s not provided, try to describe it by yourself.\n\nIf **shot size or spatial position** changes, describe how these transitions occur clearly, specifying both the **initial and final states**.\n\nIn addition, the human-written captions below already describe the subjects (if any) and the scenery in this video but do not capture their spatial composition and movements. Use these captions as a reference, but do not rely on them completely. Your goal is to fully capture the spatial framing and movements in this video. Don’t write too much about the subject's or scenery's appearance.\n\nHuman-Written Description for Subjects: **The video shows a man with short, brown hair, wearing a gray hoodie and a necklace, appearing within a circular frame that resembles a digital interface. A smaller circular frame appears, featuring a Black man in a red hoodie, with a focused expression. The setting suggests a virtual or digital communication scenario, viewed from a first-person perspective as the red hoodie man's hand briefly appears at the bottom of the screen. **\nHuman-Written Description for Scenery: **The video presents a first-person POV shot set indoors, featuring a modern, high-tech environment with floor-to-ceiling windows. A large, transparent screen displays a video call interface with a circular view of an outdoor urban setting, showing buildings and a clear sky. The time of day appears to be daytime, with bright natural light illuminating the scene. The screen also includes a smaller circular inset showing another person, suggesting a conversation. The room is sleek and futuristic, with colorful walls in shades of blue, pink, and white, and a low shelving unit on the left side, emphasizing digital communication technology.**\n\nPlease note that the video features a **complex scenario** with ambiguous subjects or it is an abstract shot. Please try your best to describe the spatial positions and movements of the main subjects or objects in the video.\n\nShot Size Information: The video features a complex scenario with ambiguous subjects or it is an abstract shot. Please try your best to describe the spatial positions and movements of the main subjects or objects in the video. Do not use shot size to describe the spatial framing.",
"review_timestamp": "2025-05-06T00:41:38.679606",
"reviewer": "Siyuan Cen",
"reviewer_double_check": true,
"timestamp": "2025-03-21T10:04:46.644403",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"negative_example": null,
"status": "approved"
},
"subject": {
"caption_data": {
"caption_rating_score": 5,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The video presents a dynamic, continuous sequence from a first-person perspective, featuring a man with light skin and dark hair displayed on a futuristic, semi-transparent screen. He is shown from the shoulders up, encased in a glowing digital bubble, wearing a grey V-neck shirt and a necklace, and speaking directly to the viewer with subtle shifts in expression. The screen rests on a cylindrical metallic base with several ports glowing in pink, orange, blue, and white, some of which have devices plugged into them. As the video progresses, a smaller circular frame appears, showcasing a Black man in a red hoodie with a focused expression. Midway through, the hand of the person filming, wearing a red sleeve, interacts with the screen's base, unplugging and replugging several USB devices. The background visible through the screen is a blurred cityscape.",
"final_feedback": "1. Consider using continuous video footage instead of relying solely on the first and second frames for a more dynamic presentation.\n2. Additionally, this is a first-person perspective, which should be noted in the description.\n3. As the video progresses, a smaller circular frame emerges, showcasing a Black man in a red hoodie with a focused expression.\n4. Midway through the video, the hand of the person filming appears below the screen, unplugging and then replugging several USB devices.",
"gpt_caption": "The video presents a dynamic, continuous sequence from a first-person perspective, featuring a man with light skin and dark hair displayed on a futuristic, semi-transparent screen. He is shown from the shoulders up, encased in a glowing digital bubble, wearing a grey V-neck shirt and a necklace, and speaking directly to the viewer with subtle shifts in expression. The screen rests on a cylindrical metallic base with several ports glowing in pink, orange, blue, and white, some of which have devices plugged into them. As the video progresses, a smaller circular frame appears, showcasing a Black man in a red hoodie with a focused expression. Midway through, the hand of the person filming, wearing a red sleeve, interacts with the screen's base, unplugging and replugging several USB devices. The background visible through the screen is a blurred cityscape.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "1. Consider using continuous video footage instead of relying solely on the first and second frames for a more dynamic presentation.\n2. Additionally, this is a first-person perspective, which should be noted in the description.\n3. As the video progresses, a smaller circular frame emerges, showcasing a Black man in a red hoodie with a focused expression.\n4. Midway through the video, the hand of the person filming appears below the screen, unplugging and then replugging several USB devices.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "1. Using video instead of the first frame and the second frame. \n2. also 这是第一人称视角 你应该描述这点\n3. as the video unfolds, a smaller circular frame appears, featuring a Black man in a red hoodie, with a focused expression.\n4. 中途拍摄者的手出现在screen下方，拔掉了几个usb又重新插上",
"original_annotator": "Yuhan Huang",
"pre_caption": "The video features a man displayed on a futuristic, semi-transparent screen. The man, with light skin and dark hair, is shown from the shoulders up, encased in a glowing digital bubble. He wears a grey V-neck shirt and a necklace, looking directly at the viewer and speaking, his expression shifting slightly throughout. The screen rests on a cylindrical metallic base equipped with several ports, some of which glow with pink, orange, blue, and white light and have devices plugged into them. A hand, belonging to an unseen person wearing a red sleeve, reaches from the bottom foreground to interact with one of the glowing ports on the screen's base. The background seen through the screen is a blurred cityscape.",
"pre_caption_llm": "gemini-2.5-pro-preview-03-25",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the subjects in this video. Keep the description concise and clear, focusing on subject types and visual attributes. Since this is the first and last frame of a video, refer to it as \"The video\" instead of mentioning \"image\" or \"frame.\" Describe the entire video by combining details from both frames, rather than describing them separately. Don't mention the background or motion unless it's necessary to distinguish subjects by location, action, or relationships. You must avoid describing what is not visible or what you are unsure about. You must use simple, natural English and ensure the description is a clear, concise, and coherent paragraph that highlights the most essential details. You must avoid subjective adjectives that convey emotions. Whenever you mention a subject, please describe its key visual attributes in detail. Return only the one-paragraph video description without Markdown formatting or introductory text.\n\nClearly identify each subject’s type, using precise terms such as “man,” “woman,” “dog,” “car,” or “tree,” rather than vague words like “thing” or “item.” If the subject type is ambiguous, use your best judgment and briefly explain your reasoning.\n\nDescribe key visual attributes with specific and descriptive language. For people, include details such as clothing color and style, skin tone, hairstyle, facial hair, age (if discernible), gender, ethnicity (if relevant and clear), and facial expression. For objects, describe their color, material, shape, and distinguishing features like texture or markings. Additionally, note the subject’s pose and orientation within the frame, such as standing, sitting, walking, or facing a certain direction. Pay attention to any objects that are not in their usual state, like a tilted lamp or an open book lying face down.\n\nIf there are multiple subjects to describe, ensure clarity in referring to each. The simplest way is by type, such as “the man,” “the dog,” or “the tree.” If multiple subjects belong to the same category, distinguish them using unique appearance traits (e.g., “the woman in the red dress,” “the man with the beard”), location within the scene (e.g., “the man on the left,” “the car in the midground”), actions (e.g., “the child playing with a ball,” “the bird flying”), or relationships to each other (e.g., “the man next to the woman,” “the first man that enters the frame”). Also, when describing multiple subjects, the order in which they are mentioned matters. Prioritize based on relevance, starting with the largest or most centered subject. If the scene unfolds over time, describe subjects in the order they appear. If temporal order isn’t relevant, begin with the most visually striking or important subject before moving to less prominent ones. The goal is to provide enough detail so that anyone reading the description can easily identify each subject.\n\nPlease note that the video features a **complex scenario** with ambiguous subjects or it is an abstract shot. Please try your best to describe the main subjects or objects in the video.\n\nPlease avoid using phrases like \"the first frame\" or \"the last frame\" in your description. Instead, refer to the entire sequence simply as \"the video.\" Your description should integrate observations from all frames into a cohesive, temporally and logically consistent narrative, rather than describing frames in isolation. Whenever you mention a subject, be sure to include detailed descriptions of its key visual attributes. The final output should be a single, fluent paragraph describing the video, with no Markdown formatting or introductory text.",
"review_timestamp": "2025-05-05T16:54:29.503522",
"reviewer": "Siyuan Cen",
"reviewer_double_check": false,
"timestamp": "2025-05-05T16:58:56.743656",
"user": "Siyuan Cen",
"workflow_type": "improved_precaption"
},
"negative_example": {
"caption_rating_score": 4,
"feedback_is_needed": true,
"feedback_rating_score": 4,
"final_caption": "The video shows a man with short, brown hair, wearing a gray hoodie and a necklace, appearing within a circular frame that resembles a digital interface. A smaller circular frame appears, featuring a Black man in a red hoodie, with a focused expression. The setting suggests a virtual or digital communication scenario, viewed from a first-person perspective as the red hoodie man's hand briefly appears at the bottom of the screen. ",
"final_feedback": "You should use video rather than the first and second frames. Additionally, this is a first-person perspective, as the videographer's hand briefly appears at the bottom of the screen. The man in the red hoodie is Black.",
"gpt_caption": "The video shows a man with short, reddish-brown hair, wearing a gray hoodie and a necklace, appearing within a circular frame that resembles a digital interface. His expression is neutral at first and becomes more engaged as the video progresses. A smaller circular frame appears, featuring a Black man in a red hoodie and a hat, with a focused expression. The setting suggests a virtual or digital communication scenario, viewed from a first-person perspective as the videographer's hand briefly appears at the bottom of the screen. The main subject is centrally positioned, with the second subject appearing as an inset.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The video should use continuous footage rather than distinct first and second frames. Additionally, this is a first-person perspective, as the videographer's hand briefly appears at the bottom of the screen. The man in the red hoodie is Black.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "Using video instead of the first frame and the second frame. also 这是第一人称，中途拍摄者的手出现在screen下方，红衣男子是黑人",
"pre_caption": "The video features a man with short, reddish-brown hair, wearing a gray hoodie and a necklace, appearing within a circular frame that resembles a digital interface. His expression is neutral in the first frame and more engaged in the second. In the second frame, a smaller circular frame appears, showing another man wearing a red hoodie and a hat, with a focused expression. The setting suggests a virtual or digital communication scenario, with the main subject centrally positioned and the second subject appearing as an inset.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the subjects in this video. Keep the description concise and clear, focusing on subject types and visual attributes. Since this is the first and last frame of a video, refer to it as \"The video\" instead of mentioning \"image\" or \"frame.\" Describe the entire video by combining details from both frames, rather than describing them separately. Don't mention the background or motion unless it's necessary to distinguish subjects by location, action, or relationships. You must avoid describing what is not visible or what you are unsure about. You must use simple, natural English and ensure the description is a clear, concise, and coherent paragraph that highlights the most essential details. You must avoid subjective adjectives that convey emotions. Return only the description.\n\nClearly identify each subject’s type, using precise terms such as “man,” “woman,” “dog,” “car,” or “tree,” rather than vague words like “thing” or “item.” If the subject type is ambiguous, use your best judgment and briefly explain your reasoning.\n\nDescribe key visual attributes with specific and descriptive language. For people, include details such as clothing color and style, hairstyle, facial hair, age (if discernible), gender, ethnicity (if relevant and clear), and facial expression. For objects, describe their color, material, shape, and distinguishing features like texture or markings. Additionally, note the subject’s pose and orientation within the frame, such as standing, sitting, walking, or facing a certain direction. Pay attention to any objects that are not in their usual state, like a tilted lamp or an open book lying face down.\n\nIf there are multiple subjects to describe, ensure clarity in referring to each. The simplest way is by type, such as “the man,” “the dog,” or “the tree.” If multiple subjects belong to the same category, distinguish them using unique appearance traits (e.g., “the woman in the red dress,” “the man with the beard”), location within the scene (e.g., “the man on the left,” “the car in the midground”), actions (e.g., “the child playing with a ball,” “the bird flying”), or relationships to each other (e.g., “the man next to the woman,” “the first man that enters the frame”). Also, when describing multiple subjects, the order in which they are mentioned matters. Prioritize based on relevance, starting with the largest or most centered subject. If the scene unfolds over time, describe subjects in the order they appear. If temporal order isn’t relevant, begin with the most visually striking or important subject before moving to less prominent ones. The goal is to provide enough detail so that anyone reading the description can easily identify each subject.\n\nPlease note that the video features a **complex scenario** with ambiguous subjects or it is an abstract shot. Please try your best to describe the main subjects or objects in the video.",
"timestamp": "2025-03-12T00:18:11.366486",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"status": "rejected"
}
}
|
-2uIa-XMJC0.8.0.mp4
|
https://huggingface.co/datasets/zhiqiulin/video_captioning/resolve/main/-2uIa-XMJC0.8.0.mp4
|
{
"camera": {
"caption_data": {
"caption_rating_score": null,
"feedback_is_needed": false,
"feedback_rating_score": null,
"final_caption": "The video features mild barrel distortion, causing straight lines near the edges to bow outward. The camera is positioned at eye level with a level angle, maintaining a deep focus and large depth of field. It simulates an FPV view within a room, continuously panning and tilting to explore the surroundings. The movement is slightly unsteady with some shaking, intensifying toward the end as the camera shakes intensely when a person punches toward the viewer.",
"final_feedback": null,
"gpt_caption": null,
"gpt_caption_llm": null,
"gpt_caption_prompt": null,
"gpt_feedback": null,
"gpt_feedback_llm": null,
"gpt_feedback_prompt": null,
"initial_caption_rating_score": 5,
"initial_feedback": null,
"original_annotator": null,
"pre_caption": "The video features mild barrel distortion, causing straight lines near the edges to bow outward. The camera is positioned at eye level with a level angle, maintaining a deep focus and large depth of field. It simulates an FPV view within a room, continuously panning and tilting to explore the surroundings. The movement is slightly unsteady with some shaking, intensifying toward the end as the camera shakes intensely when a person punches toward the viewer.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the **video’s and camera’s configuration**, covering **video speed, lens distortion, camera angle, camera height, movements (translation, rotation, steadiness, intensity, complexity), and focus (depth, focus plane, focus changes).** \n\nIf **video speed** is altered, specify the type, such as *time-lapse* (“Clouds move rapidly across the sky”), *slow-motion*, or *speed ramp* (a mix of fast and slow speeds). If the video is *time-reversed* or *stop-motion*, note this as well. \n\nIf **lens distortion** is present, describe the type and degree. For example, *fisheye distortion* creates extreme curvature, while *barrel distortion* causes mild outward bowing of straight lines near the edges. \n\nDescribe the **camera height** in relation to the ground, such as *eye-level, ground-level, overhead-level, or underwater*. If height changes due to movement, mention how it transitions. Similarly, specify the **camera angle**, such as *bird’s eye, high angle, low angle, or worm’s eye*, noting any shifts within the video. If a **Dutch angle** (tilted horizon) is present, indicate whether it remains fixed or varies due to camera rolling. \n\nIf discernible, describe the **camera focus and depth of field**. For example, *deep focus* keeps all elements sharp, while *shallow depth of field* blurs the background or foreground. If focus changes dynamically, note whether it’s a *rack focus* (shifting focus between subjects) or *focus tracking* (following a subject’s depth movement). If the video lacks realistic depth of field, describe whether it appears artificial, overly blurry, or lacking in cinematic focus effects. \n\nIf the **camera is static**, no further description is needed. If it moves, describe its **type, direction, and speed**. Specify movements such as *tracking (following a subject), arcing (circling around), dollying (moving forward/backward), trucking (moving left/right), panning (rotating horizontally), tilting (angling vertically), pedestaling (lifting/lowering), or rolling (rotating around the lens axis).* If motion intensity varies, distinguish between slow and rapid movements and present them in **temporal order** (e.g., *“The camera first pans right, then tilts upward to follow the subject”*). If motion appears **fragmented or random**, summarize it concisely rather than over-describing each small movement (e.g., *“The first-person camera moves erratically, frequently changing direction without a clear pattern”*).\n\nCrucially, instead of inferring these attributes from the video, we have already provided human-labeled ground truth for some of the elements specified above. You should directly use this information in your description and should not infer any details that are not already provided. Your description should be brief, and if anything is normal or unremarkable, you do not need to include it (e.g., if the video is at regular playback speed, there is no need to mention it).\n\n**Playback Speed:** A regular-speed video where events play in real-time without acceleration or slowing. (no need to mention).\n\n**Lens Distortion:** The video features mild barrel distortion causing straight lines near the edges to bow outward.\n\n**Camera Height:** The camera is at an eye level (above the waist).\n\n**Camera Angle:** The camera angle is at a level angle (looking straight ahead).\n\n**Camera Focus:** The camera uses a deep focus with a large depth of field.\n\n**Camera Motion:** The camera simulates an FPV view within a room, continuously panning and tilting to explore the surroundings. Toward the end, it shakes intensely as a person punches toward the viewer.\n\n**Camera Steadiness:** The camera movement is slightly unsteady with some shaking.",
"review_timestamp": "2025-05-29T02:10:09.872471",
"reviewer": "Shihang Zhu",
"reviewer_double_check": true,
"timestamp": "2025-03-21T10:29:37.072308",
"user": "Yuhan Huang",
"workflow_type": "perfect_precaption"
},
"negative_example": null,
"status": "approved"
},
"motion": {
"caption_data": {
"caption_rating_score": 5,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The video begins with a blurred, dark room filled with multiple screens displaying various images and data. As the camera moves, the screens continue to show different, colorful displays, including brain scans and other data visualizations. A man appears in the background, walking towards the camera. He gets closer, and his face becomes clearer, showing a focused expression. The video concludes with the man punching the camera, with the screens still visible in the background.",
"final_feedback": "After the man approaches the camera, he punches it. Additionally, the phrases \"and the camera captures his movement in detail\" and \"The video ends with the man standing close to the camera, with the screens still visible in the background\" should be removed, as they do not accurately describe the video content.",
"gpt_caption": "The video begins with a blurred, dark room filled with multiple screens displaying various images and data. As the camera moves, the screens continue to show different, colorful displays, including brain scans and other data visualizations. A man appears in the background, walking towards the camera. He gets closer, and his face becomes clearer, showing a focused expression. The video concludes with the man punching the camera, with the screens still visible in the background.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "After the man approaches the camera, he punches it. Additionally, the phrases \"and the camera captures his movement in detail\" and \"The video ends with the man standing close to the camera, with the screens still visible in the background\" should be removed, as they do not accurately describe the video content.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "男子走近相机后，给了相机一拳头. And you should delete \"and the camera captures his movement in detail. The video ends with the man standing close to the camera, with the screens still visible in the background\" because this is not accurate.",
"original_annotator": "Yuhan Huang",
"pre_caption": "The video begins with a blurred, dark room filled with multiple screens displaying various images and data. As the camera moves, the screens continue to show different, colorful displays, including brain scans and other data visualizations. A man appears in the background, walking towards the camera. He gets closer, and his face becomes clearer, showing a focused expression. The man continues to approach, and the camera captures his movement in detail. The video ends with the man standing close to the camera, with the screens still visible in the background.",
"pre_caption_llm": "tarsier-recap-7b",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the subject’s motion in this video, ensuring actions are presented in **chronological order** if multiple movements occur (e.g., \"The bird first takes flight, then soars in a circle, and finally lands on a branch\"). Focus on the subject's motion rather than repeating details already included in the human-written subject descriptions. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B.' Return only the one-paragraph video description without Markdown formatting or introductory text.\n\nIf the subject in the video has no movement, please briefly mention that without going into too much detail.\n\nPlease only describe the content of the video. Don't mention the details of the subject's appearance unless you need to differentiate between multiple subjects by their appearance. Clearly describe the subject's motion.\n\nAvoid abstract descriptions, such as \"The car maintains a low, sleek profile as it maneuvers the bend, emphasizing its speed and agility\" and \"emphasizing its speed and agility as it maneuvers through the turn.\"\n\nBelow are detailed instructions:\n\nDescribe **individual subject actions** with clarity, specifying how they move rather than using generic descriptions. For example, instead of “a person is running,” say “a runner sprints across the finish line.”\n\nIf the subject interacts with an **object**, specify the type of interaction and its effect. Instead of “a person is working,” say “a construction worker operates a jackhammer, breaking up the pavement.”\n\nIf there are **interactions between subjects**, describe the nature of their relationship and movements relative to each other. Instead of “people are fighting,” say “two boxers exchange blows in the ring, circling each other cautiously.”\n\nIf there are collective behaviors for a group of subjects, describe **group activities** with specificity. Instead of “birds are flying,” say “a flock of geese flies in a V-formation across the horizon.” Instead of “people are walking,” say “a crowd of protesters marches down the street, carrying signs and banners.” Clearly convey the type of group, their coordinated actions, and any notable patterns in their movement.\n\nIn addition, the human-written caption below already describes the subjects (if any) in this video but does not comprehensively capture their motion. Use this caption as a reference or starting point to draft the description, but do not rely on it completely. Expand or refine it to fully capture the subjects' motion and dynamics. However, if the caption does not mention any subjects, do not add any description—simply note this briefly in the description. Please note that this caption is only to help you clarify which subject's motion needs to be analyzed, rather than adding more content based on this description.\n\nHuman-Written Caption: **The video shows a white man in a room filled with screens. He is wearing black pants, a gray hoodie, a necklace, and has short hair. The focus is on his interaction with the camera.**",
"review_timestamp": "2025-05-07T02:30:52.293300",
"reviewer": "Siyuan Cen",
"reviewer_double_check": false,
"timestamp": "2025-05-07T02:35:41.516016",
"user": "Siyuan Cen",
"workflow_type": "improved_precaption"
},
"negative_example": {
"caption_rating_score": 4,
"feedback_is_needed": true,
"feedback_rating_score": 5,
"final_caption": "The video begins with a blurred, dark room filled with multiple screens displaying various images and data. As the camera moves, the screens continue to show different, colorful displays, including brain scans and other data visualizations. A man appears in the background, walking towards the camera. He gets closer, and his face becomes clearer. As he reaches the camera, he suddenly punches it.",
"final_feedback": "After the man approaches the camera, he punches it.",
"gpt_caption": "The video begins with a blurred, dark room filled with multiple screens displaying various images and data. As the camera moves, the screens continue to show different, colorful displays, including brain scans and other data visualizations. A man appears in the background, walking towards the camera. He gets closer, and his face becomes clearer, showing a focused expression. The man continues to approach, and the camera captures his movement in detail. As he reaches the camera, he suddenly punches it, and the video ends with the screens still visible in the background.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "After the man approaches the camera, he punches it.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "男子走近相机后，给了相机一拳头",
"pre_caption": "The video begins with a blurred, dark room filled with multiple screens displaying various images and data. As the camera moves, the screens continue to show different, colorful displays, including brain scans and other data visualizations. A man appears in the background, walking towards the camera. He gets closer, and his face becomes clearer, showing a focused expression. The man continues to approach, and the camera captures his movement in detail. The video ends with the man standing close to the camera, with the screens still visible in the background.",
"pre_caption_llm": "tarsier-recap-7b",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Provide a concise yet informative description of the subject’s motion in this video, ensuring actions are presented in **chronological order** if multiple movements occur (e.g., \"The bird first takes flight, then soars in a circle, and finally lands on a branch\"). Since this represents the first and last frames of a video, refer to it as 'The video' instead of mentioning 'image' or 'frame.' Describe the entire video by combining details from both frames rather than separately. Focus on the setting and scenery rather than detailed subject descriptions. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B.' Return only the description.\n\nIf the subject in the video has no movement, please briefly mention that without going into too much detail.\n\nPlease only describe the content of the video. Don't mention the details of the subject's appearance unless you need to differentiate between multiple subjects by their appearance. Clearly describe the subject's motion.\n\nAvoid abstract descriptions, such as \"The car maintains a low, sleek profile as it maneuvers the bend, emphasizing its speed and agility\" and \"emphasizing its speed and agility as it maneuvers through the turn.\"\n\nBelow are detailed instructions:\n\nDescribe **individual subject actions** with clarity, specifying how they move rather than using generic descriptions. For example, instead of “a person is running,” say “a runner sprints across the finish line.”\n\nIf the subject interacts with an **object**, specify the type of interaction and its effect. Instead of “a person is working,” say “a construction worker operates a jackhammer, breaking up the pavement.”\n\nIf there are **interactions between subjects**, describe the nature of their relationship and movements relative to each other. Instead of “people are fighting,” say “two boxers exchange blows in the ring, circling each other cautiously.”\n\nIf there are collective behaviors for a group of subjects, describe **group activities** with specificity. Instead of “birds are flying,” say “a flock of geese flies in a V-formation across the horizon.” Instead of “people are walking,” say “a crowd of protesters marches down the street, carrying signs and banners.” Clearly convey the type of group, their coordinated actions, and any notable patterns in their movement.\n\nIn addition, the human-written caption below already describes the subjects (if any) in this video but does not comprehensively capture their motion. Use this caption as a reference or starting point to draft the description, but do not rely on it completely. Expand or refine it to fully capture the subjects' motion and dynamics. However, if the caption does not mention any subjects, do not add any description—simply note this briefly in the description. Please note that this caption is only to help you clarify which subject's motion needs to be analyzed, rather than adding more content based on this description.\n\nHuman-Written Caption: **The video shows a white man in a room filled with screens. He is wearing black pants, a gray hoodie, a necklace, and has short hair. The focus is on his interaction with the camera.**",
"timestamp": "2025-03-19T10:57:18.138068",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"status": "rejected"
},
"scene": {
"caption_data": {
"caption_rating_score": null,
"feedback_is_needed": false,
"feedback_rating_score": null,
"final_caption": "The video presents a first-person POV shot set in a dimly lit indoor environment, characterized by a futuristic and abstract aesthetic. The scene is dominated by a variety of glowing, neon-colored elements that create a surreal atmosphere. The lighting is primarily blue and purple, casting an otherworldly glow over the surroundings. The setting appears to be filled with digital or holographic displays, contributing to a high-tech ambiance. The overall mood is enhanced by the dynamic interplay of light and shadow, with the neon elements providing the primary source of illumination. The scene transitions smoothly, maintaining a consistent color scheme and visual style throughout.",
"final_feedback": null,
"gpt_caption": null,
"gpt_caption_llm": null,
"gpt_caption_prompt": null,
"gpt_feedback": null,
"gpt_feedback_llm": null,
"gpt_feedback_prompt": null,
"initial_caption_rating_score": 5,
"initial_feedback": null,
"original_annotator": null,
"pre_caption": "The video presents a first-person POV shot set in a dimly lit indoor environment, characterized by a futuristic and abstract aesthetic. The scene is dominated by a variety of glowing, neon-colored elements that create a surreal atmosphere. The lighting is primarily blue and purple, casting an otherworldly glow over the surroundings. The setting appears to be filled with digital or holographic displays, contributing to a high-tech ambiance. The overall mood is enhanced by the dynamic interplay of light and shadow, with the neon elements providing the primary source of illumination. The scene transitions smoothly, maintaining a consistent color scheme and visual style throughout.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the overall scene, including the point of view, environment, setting, time of day, and notable visual elements. If subjects are present, ensure their placement and context complement the scene without excessive detail. Since this represents the first and last frames of a video, refer to it as 'The video' instead of mentioning 'image' or 'frame.' Describe the entire video by combining details from both frames rather than separately. Focus on the setting and scenery rather than detailed subject descriptions. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B.' Return only the description.\n\nIf relevant, indicate the **point of view**, such as first-person, drone shot, or dashcam, and describe how it influences the viewer’s perception. Specify the **setting** by clearly identifying whether it is indoors or outdoors, using precise language. If the location is known, state it explicitly (e.g., \"Times Square\" or \"Tokyo subway station\"). Otherwise, describe defining features such as “a narrow alley with graffiti-covered walls” or “a vast desert with rolling dunes.” Mention the **time of day** and any notable **architectural or natural features**, such as buildings, roads, forests, or bodies of water. Include relevant **weather conditions** if applicable, like “a rainy street with wet pavement reflecting city lights” or “a snowy mountain pass covered in thick fog.” For indoor settings, describe key **furniture or props** that establish the environment, such as “a wooden desk cluttered with books and a vintage lamp.” If notable, mention the **style** of the scene, such as a monochromatic color scheme or a vibrant carnival with neon lights. \n\nIf the scene involves **motion or changes**, describe natural elements like wind blowing through trees or waves crashing against the shore, as well as human-made movements such as traffic flowing on a highway or pedestrians walking along a busy street. If **scene transitions** occur, explain how they happen, whether through changes in time (e.g., “The shot transitions from day to night”) or movement-based shifts (e.g., “The camera pans to reveal an alley behind the main street”). Use precise language to refer to different elements within the scene, prioritizing the most prominent details and organizing descriptions logically from the overall setting to more specific features.\n\nIn addition, you do not need to infer the camera's point of view, as this information is already provided. Please integrate the following point of view information into your caption:\n\nPoint of View Information: \"**A first-person POV shot.**\"",
"review_timestamp": "2025-05-21T06:58:11.914438",
"reviewer": "Shihang Zhu",
"reviewer_double_check": true,
"timestamp": "2025-03-19T05:12:49.176710",
"user": "Yuhan Huang",
"workflow_type": "perfect_precaption"
},
"negative_example": null,
"status": "approved"
},
"spatial": {
"caption_data": {
"caption_rating_score": 4,
"feedback_is_needed": true,
"feedback_rating_score": 4,
"final_caption": "The video begins with a wide shot of a dimly lit room filled with screens, creating a futuristic ambiance. Initially, there is no primary subject, and the focus remains on the abstract, neon-colored elements in the background. As the video progresses, a man enters from the left side in the background, moving slightly right of center in the middle ground as he approaches the camera. The shot transitions from a full shot of the man to a medium shot as he interacts with the camera by punching it. The neon elements persist in the background, reinforcing the high-tech atmosphere. After the interaction, the shot returns to a full shot of the man, maintaining the room's immersive and surreal environment.",
"final_feedback": "The video initially presents a wide shot of the room. The person enters from the left side in the background and moves to a position slightly right of center, rather than directly in the center of the frame.",
"gpt_caption": "The video begins with a wide shot of a dimly lit room filled with screens, creating a futuristic ambiance. Initially, there is no major subject, and the focus is on the abstract, neon-colored elements in the background. As the video progresses, a person enters the scene from the left side in the background, moving toward the camera, which is positioned at the subject's height. The person moves to a position slightly right of center as they approach and interact with the camera by punching it. The shot transitions from a full shot to a middle shot during this interaction. The neon elements remain in the background, providing a consistent high-tech atmosphere. After the interaction, the shot returns to a full shot, maintaining the room's immersive and surreal environment.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "The video initially presents a wide shot of the room. The person enters from the left side in the background and moves to a position slightly right of center, rather than directly in the middle of the frame.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "一开始是这个房间的wide shot。the person enters from the left in the background. 人从左边走到画面中间偏右的位置，不在正中间。",
"original_annotator": null,
"pre_caption": "The video begins with a dimly lit room filled with screens, creating a futuristic ambiance. Initially, there is no major subject, and the focus is on the abstract, neon-colored elements in the background. As the video progresses, a person enters the scene from the background, moving toward the camera, which is positioned at the subject's height. The shot transitions from a full shot to a middle shot as the person approaches and interacts with the camera by punching it. The subject is centrally positioned, moving from the background to the foreground. The neon elements remain in the background, providing a consistent high-tech atmosphere. After the interaction, the shot returns to a full shot, maintaining the room's immersive and surreal environment.",
"pre_caption_llm": "qwen2.5-vl-72b",
"pre_caption_mode": "Video",
"pre_caption_prompt": "Analyze the subjects and elements in this video and provide a concise yet informative description of how subjects and elements are framed within the scene, including their **shot size, position, and changes** within the frame. Your goal is to describe the **spatial framing and dynamics** within the shot. Avoid describing anything not visible or uncertain. Use simple, natural English to create a clear, concise, and coherent paragraph that highlights essential details. Avoid emotional or subjective adjectives. Avoid speculative statements like 'there might be,' 'it appears,' or ambiguous options like 'A or B'. Return only the description.\n\nFirst, specify the **shot size** based on the subject's size in the frame if major subjects are present. If the shot size is unclear, describe how much of the subject is visible. If no major subject exists (e.g., a scenery shot), describe the shot size in relation to the scenery.\n\nNext, describe the **spatial position of subjects and elements in the video**, if relevant. Indicate their approximate **2D position** within the frame using terms like **left, right, bottom left, bottom right, top right, top left, bottom, top, or center**. Additionally, describe their **3D position** within the scene as **foreground, middle ground, or background**. Analyze as many elements as possible, and for each element mentioned, provide both its **2D and 3D position**.\n\nFinally, describe the **camera’s height relative to the subject**, if relevant. Indicate whether the camera is positioned at the subject's height, above them, or below them. We already have this information provided at the end. If it’s not provided, try to describe it by yourself.\n\nIf **shot size or spatial position** changes, describe how these transitions occur clearly, specifying both the **initial and final states**.\n\nIn addition, the human-written captions below already describe the subjects (if any) and the scenery in this video but do not capture their spatial composition and movements. Use these captions as a reference, but do not rely on them completely. Your goal is to fully capture the spatial framing and movements in this video. Don’t write too much about the subject's or scenery's appearance.\n\nHuman-Written Description for Subjects: **The video shows a white man in a room filled with screens. He is wearing black pants, a gray hoodie, a necklace, and has short hair. The focus is on his interaction with the camera.**\nHuman-Written Description for Scenery: **The video presents a first-person POV shot set in a dimly lit indoor environment, characterized by a futuristic and abstract aesthetic. The scene is dominated by a variety of glowing, neon-colored elements that create a surreal atmosphere. The lighting is primarily blue and purple, casting an otherworldly glow over the surroundings. The setting appears to be filled with digital or holographic displays, contributing to a high-tech ambiance. The overall mood is enhanced by the dynamic interplay of light and shadow, with the neon elements providing the primary source of illumination. The scene transitions smoothly, maintaining a consistent color scheme and visual style throughout.**\n\nThe description below already mentions the spatial framing information about the subjects or scenery in this video. Use this caption as a reference to draft the spatial framing and dynamics description. Simply expand on it to fully capture other spatial positions and movements. Do not infer the any spatial framing information already mentioned below.\n\nShot Size Information: The video takes place in a room filled with screens. A person walks toward the camera and then punches it. The shot size transitions from no subject at the beginning to a full shot when the person appears, then to a middle shot as he punch the camera, and finally back to a full shot.\n\nCamera Height Relative to Subjects: There is no subject at the beginning. Once the subject appears, the camera is at the same height as the subject.",
"review_timestamp": "2025-05-06T00:44:37.730003",
"reviewer": "Siyuan Cen",
"reviewer_double_check": true,
"timestamp": "2025-04-02T10:37:56.899214",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"negative_example": null,
"status": "approved"
},
"subject": {
"caption_data": {
"caption_rating_score": 5,
"feedback_is_needed": true,
"feedback_rating_score": 4,
"final_caption": "The video shows a white man in a room filled with screens. He is wearing black pants, a gray hoodie, a necklace, and has short hair. The focus is on his interaction with the camera.",
"final_feedback": "Please avoid describing the detailed motions and shot sizes. The individual in the video is a white man wearing black pants, a gray hoodie, a necklace, and has short hair.",
"gpt_caption": "The video shows a white man in a room filled with screens. He is wearing black pants, a gray hoodie, a necklace, and has short hair. The focus is on his interaction with the camera.",
"gpt_caption_llm": "gpt-4o-2024-08-06",
"gpt_caption_prompt": "Given a video caption and user feedback, please provide an improved version of the caption that addresses the feedback. Note that the user feedback could be poorly written, so please try your best to guess what it means.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with the improved caption only, without quotation marks or JSON formatting.",
"gpt_feedback": "Please avoid describing the detailed motions and shot sizes. The individual in the video is a Caucasian wearing black pants, a gray hoodie, a necklace, and has short hair.",
"gpt_feedback_llm": "gpt-4o-2024-08-06",
"gpt_feedback_prompt": "Given a video caption and user feedback, rewrite the feedback to enhance clarity, professionalism, and completeness while preserving its original intent. If the feedback is vague, overly brief, or poorly written, infer the likely meaning and refine it accordingly.\n\nOriginal caption: {pre_caption}\nUser feedback: {feedback}\n\nRespond with only the polished feedback, without JSON formatting or additional commentary.",
"initial_caption_rating_score": 3,
"initial_feedback": "不用描述太细节的motion和shot size，这个白人穿黑色裤子和灰色连帽衫，带项链，短发",
"original_annotator": null,
"pre_caption": "The video features a person in a room filled with screens. Initially, the scene is devoid of any subjects, but the person soon walks toward the camera. As they approach, the shot transitions to a full view of the person, who is wearing dark clothing. The person then punches the camera, shifting the shot to a middle view. After the action, the shot returns to a full view, with the person still visible among the screens. The focus is on the person's movement and interaction with the camera.",
"pre_caption_llm": "gpt-4o-2024-08-06",
"pre_caption_mode": "Image (First-and-Last-Frame)",
"pre_caption_prompt": "Provide a concise yet informative description of the subjects in this video. Keep the description concise and clear, focusing on subject types and visual attributes. Since this is the first and last frame of a video, refer to it as \"The video\" instead of mentioning \"image\" or \"frame.\" Describe the entire video by combining details from both frames, rather than describing them separately. Don't mention the background or motion unless it's necessary to distinguish subjects by location, action, or relationships. You must avoid describing what is not visible or what you are unsure about. You must use simple, natural English and ensure the description is a clear, concise, and coherent paragraph that highlights the most essential details. You must avoid subjective adjectives that convey emotions. Return only the description.\n\nClearly identify each subject’s type, using precise terms such as “man,” “woman,” “dog,” “car,” or “tree,” rather than vague words like “thing” or “item.” If the subject type is ambiguous, use your best judgment and briefly explain your reasoning.\n\nDescribe key visual attributes with specific and descriptive language. For people, include details such as clothing color and style, hairstyle, facial hair, age (if discernible), gender, ethnicity (if relevant and clear), and facial expression. For objects, describe their color, material, shape, and distinguishing features like texture or markings. Additionally, note the subject’s pose and orientation within the frame, such as standing, sitting, walking, or facing a certain direction. Pay attention to any objects that are not in their usual state, like a tilted lamp or an open book lying face down.\n\nIf there are multiple subjects to describe, ensure clarity in referring to each. The simplest way is by type, such as “the man,” “the dog,” or “the tree.” If multiple subjects belong to the same category, distinguish them using unique appearance traits (e.g., “the woman in the red dress,” “the man with the beard”), location within the scene (e.g., “the man on the left,” “the car in the midground”), actions (e.g., “the child playing with a ball,” “the bird flying”), or relationships to each other (e.g., “the man next to the woman,” “the first man that enters the frame”). Also, when describing multiple subjects, the order in which they are mentioned matters. Prioritize based on relevance, starting with the largest or most centered subject. If the scene unfolds over time, describe subjects in the order they appear. If temporal order isn’t relevant, begin with the most visually striking or important subject before moving to less prominent ones. The goal is to provide enough detail so that anyone reading the description can easily identify each subject.\n\nIn addition, the human-written caption below already describes the subjects (if any) in this video, including framing information like shot size. Use this caption as a reference to draft the subject description, but do not rely on it completely. Expand or refine it to fully capture the subject’s type, attributes, appearance, unique features, pose, orientation, relationships between subjects, or any changes in the main focused subject, such as revealing, disappearing, or shifting focus. However, if the description below does not mention any subjects, please do not describe subjects and only specify the type of shot (e.g., a landscape or cityscape scenery shot or a FPV shot) and explain why there is no main subject.\n\nHuman-Written Caption: **The video takes place in a room filled with screens. A person walks toward the camera and then punches it. The shot size transitions from no subject at the beginning to a full shot when the person appears, then to a middle shot as he punch the camera, and finally back to a full shot.**",
"review_timestamp": "2025-05-06T00:00:37.155789",
"reviewer": "Siyuan Cen",
"reviewer_double_check": true,
"timestamp": "2025-03-12T00:27:01.571334",
"user": "Yuhan Huang",
"workflow_type": "improved_precaption"
},
"negative_example": null,
"status": "approved"
}
}
|
-F9hoWSAX5w.0.1.mp4
|
https://huggingface.co/datasets/zhiqiulin/video_captioning/resolve/main/-F9hoWSAX5w.0.1.mp4
| {"camera":{"caption_data":{"caption_rating_score":null,"feedback_is_needed":false,"feedback_rating_s(...TRUNCATED)
|
-F9hoWSAX5w.0.2.mp4
|
https://huggingface.co/datasets/zhiqiulin/video_captioning/resolve/main/-F9hoWSAX5w.0.2.mp4
| {"camera":{"caption_data":{"caption_rating_score":null,"feedback_is_needed":false,"feedback_rating_s(...TRUNCATED)
|
-F9hoWSAX5w.0.5.mp4
|
https://huggingface.co/datasets/zhiqiulin/video_captioning/resolve/main/-F9hoWSAX5w.0.5.mp4
| {"camera":{"caption_data":{"caption_rating_score":null,"feedback_is_needed":false,"feedback_rating_s(...TRUNCATED)
|
-F9hoWSAX5w.0.6.mp4
|
https://huggingface.co/datasets/zhiqiulin/video_captioning/resolve/main/-F9hoWSAX5w.0.6.mp4
| {"camera":{"caption_data":{"caption_rating_score":5,"feedback_is_needed":true,"feedback_rating_score(...TRUNCATED)
|
0096bcca-27a61cb3.0.mp4
|
https://huggingface.co/datasets/zhiqiulin/video_captioning/resolve/main/0096bcca-27a61cb3.0.mp4
| {"camera":{"caption_data":{"caption_rating_score":null,"feedback_is_needed":false,"feedback_rating_s(...TRUNCATED)
|
00c44d89f1c7e739b878c96864297e69c18b798305ce8b662c601e2133a03e2e.0.mp4
|
https://huggingface.co/datasets/zhiqiulin/video_captioning/resolve/main/00c44d89f1c7e739b878c96864297e69c18b798305ce8b662c601e2133a03e2e.0.mp4
| {"camera":{"caption_data":{"caption_rating_score":null,"feedback_is_needed":false,"feedback_rating_s(...TRUNCATED)
|
End of preview.
README.md exists but content is empty.
- Downloads last month
- 47