parkseg-to-fo.py

"""
Simple script to create a FiftyOne dataset from the parkseg12k dataset with NDVI calculation.
"""

# Import necessary libraries
import fiftyone as fo
import os
import numpy as np
from PIL import Image
from datasets import load_dataset

def main():
    # Create a new FiftyOne dataset
    name = "parkseg12k_train"
    dataset = fo.Dataset(name, overwrite=True, persistent=True)

    # Load the Hugging Face dataset
    try:
        hf_dataset = load_dataset("file://" + os.path.join(os.getcwd(), "parkseg12k_dataset"))
        print("Loaded dataset from local storage")
    except:
        print("Loading from HuggingFace")
        hf_dataset = load_dataset("UTEL-UIUC/parkseg12k")

    # Create directories for storing images
    images_dir = os.path.join(os.getcwd(), "parkseg12k_images_train")
    ndvi_dir = os.path.join(os.getcwd(), "parkseg12k_ndvi")
    os.makedirs(images_dir, exist_ok=True)
    os.makedirs(ndvi_dir, exist_ok=True)

    # Only process the train split
    split = "train"
    print(f"Processing {split} split...")
    samples = []

    # Process each sample
    for i, sample in enumerate(hf_dataset[split]):
        if i % 100 == 0:
            print(f"Processing sample {i}/{len(hf_dataset[split])}")
        
        # Create paths for saving images
        rgb_path = os.path.join(images_dir, f"{i}_rgb.png")
        mask_path = os.path.join(images_dir, f"{i}_mask.png")
        nir_path = os.path.join(images_dir, f"{i}_nir.png")
        ndvi_path = os.path.join(ndvi_dir, f"{i}_ndvi.npy")
        
        # Save images to disk
        sample['rgb'].save(rgb_path)
        sample['mask'].save(mask_path)
        sample['nir'].save(nir_path)
        
        # Calculate NDVI
        rgb_array = np.array(sample['rgb'])
        nir_array = np.array(sample['nir'])
        
        # Extract red channel and normalize
        red = rgb_array[:, :, 0].astype(np.float32) / 255.0
        nir = nir_array.astype(np.float32) / 255.0
        
        # Calculate NDVI: (NIR - Red) / (NIR + Red)
        numerator = nir - red
        denominator = nir + red
        
        # Avoid division by zero
        ndvi = np.where(denominator != 0, 
                         numerator / denominator,
                         0)
        
        # Clip to valid NDVI range
        ndvi = np.clip(ndvi, -1, 1)
        
        # Save NDVI array
        np.save(ndvi_path, ndvi)
        
        # Create FiftyOne sample
        fo_sample = fo.Sample(filepath=rgb_path)
        
        # Add mask as Segmentation using mask_path
        fo_sample["segmentation"] = fo.Segmentation(mask_path=mask_path)
        
        # Add NIR as Heatmap using map_path with range [0,1]
        fo_sample["nir"] = fo.Heatmap(map_path=nir_path, range=[0, 1])
        
        # Add NDVI as Heatmap with the array directly
        fo_sample["ndvi"] = fo.Heatmap(map=ndvi, range=[-1, 1])
        
        # Optional: Add NDVI statistics as metadata
        fo_sample["ndvi_mean"] = float(np.mean(ndvi))
        fo_sample["ndvi_std"] = float(np.std(ndvi))
        fo_sample["ndvi_min"] = float(np.min(ndvi))
        fo_sample["ndvi_max"] = float(np.max(ndvi))
        
        # Add to samples list
        samples.append(fo_sample)

    # Add all samples at once
    print(f"Adding {len(samples)} samples to dataset...")
    dataset.add_samples(samples)

    # Compute metadata and add dynamic fields
    dataset.compute_metadata()
    dataset.add_dynamic_sample_fields()

    return dataset

if __name__ == "__main__":
    dataset = main()