ObjectPose9D Dataset

Data Structure

Each data sample contains 4 attributes:

• source: Source dataset name (e.g., "cityscapes")
• prompt: Text description or prompt
• image: Image data (stored as binary)
• map: CNOCS Map (stored as binary in EXR format)

Cityscapes Subset

Due to license restrictions, images from the Cityscapes source cannot be redistributed directly.

• For Cityscapes samples, the image field contains only the relative path within leftImg8bit
• You must download the original Cityscapes dataset from: https://www.cityscapes-dataset.com/downloads/
• For other sources, image contains the actual binary image data

CNOCS Map

The map field stores the CNOCS Map from the paper in EXR format as binary data.

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("FudanCVL/ObjectPose9D", streaming=True)

for i, item in enumerate(dataset["train"]):
    if item["source"] == "cityscapes":
        # For cityscapes: image field is a path string
        image_path = item["image"].decode("utf-8")
    else:
        with open(f"{i}_{item['source']}.jpg", "wb") as f:
            f.write(item["image"])
    
    with open(f"{i}_{item['source']}.exr", "wb") as f:
        f.write(item["map"])

Visualizing CNOCS Maps

import numpy as np
from openexr_numpy import imread
from PIL import Image

# Read EXR map
cnocs_map = imread("map.exr")
cnocs_map_uint8 = (cnocs_map * 255).clip(0, 255).astype(np.uint8)
img = Image.fromarray(cnocs_map_uint8)
img.save("map.png")