extract_frames.py

"""Extract reference keyframes from videos using annotation metadata.

This script reads annotation JSONs to find the keyframe index for each video,
extracts that frame, and saves it as 'original.png' in the expected directory
structure for CustomDiT training.

Usage:
    python extract_frames.py \
        --video_dir ./data/videos \
        --annotation_dir ./data/annotations \
        --output_dir ./data/reference_images \
        --metadata_csv ./data/metadata/train.csv

The output directory structure will match what CustomDiT training expects:
    {output_dir}/{videoid}/original.png

For example:
    data/reference_images/pexels/videos-popular/10000201/original.png
"""

import os
import json
import argparse
from PIL import Image

try:
    import decord
    decord.bridge.set_bridge("native")
except ImportError:
    print("Error: decord is required. Install with: pip install decord")
    exit(1)


def parse_args():
    parser = argparse.ArgumentParser(description="Extract reference keyframes from videos")
    parser.add_argument("--video_dir", type=str, required=True,
                        help="Root directory containing videos (e.g., ./data/videos)")
    parser.add_argument("--annotation_dir", type=str, required=True,
                        help="Root directory containing annotation JSONs (e.g., ./data/annotations)")
    parser.add_argument("--output_dir", type=str, required=True,
                        help="Output directory for reference images (e.g., ./data/reference_images)")
    parser.add_argument("--metadata_csv", type=str, required=True,
                        help="Path to train.csv or val.csv to get list of videoids")
    parser.add_argument("--num_workers", type=int, default=4,
                        help="Number of parallel workers")
    return parser.parse_args()


def extract_keyframe(videoid, video_dir, annotation_dir, output_dir):
    """Extract keyframe for a single video."""
    basename = os.path.basename(videoid)
    output_path = os.path.join(output_dir, videoid, "original.png")

    # Skip if already extracted
    if os.path.exists(output_path):
        return "skip"

    # Read annotation JSON to get keyframe_index
    json_path = os.path.join(annotation_dir, videoid, f"{basename}.json")
    if not os.path.exists(json_path):
        return "missing_json"

    with open(json_path, "r") as f:
        data = json.load(f)

    keyframe_index = data.get("keyframe_index")
    if keyframe_index is None:
        return "no_keyframe_index"

    # Find video file
    video_path = os.path.join(video_dir, f"{videoid}.mp4")
    if not os.path.exists(video_path):
        # Try without extension (in case videoid already has extension)
        video_path = os.path.join(video_dir, videoid)
        if not os.path.exists(video_path):
            return "missing_video"

    # Extract frame
    try:
        vr = decord.VideoReader(video_path)
        # Clamp keyframe_index to valid range
        idx = min(keyframe_index, len(vr) - 1)
        frame = vr[idx]
        img = Image.fromarray(frame.asnumpy()).convert("RGB")

        os.makedirs(os.path.dirname(output_path), exist_ok=True)
        img.save(output_path)
        return "ok"
    except Exception as e:
        return f"error: {e}"


def main():
    args = parse_args()

    # Get unique videoids from metadata CSV
    videoids = set()
    with open(args.metadata_csv, "r") as f:
        header = f.readline()  # skip header
        for line in f:
            parts = line.strip().split(",", 2)
            if parts:
                videoids.add(parts[0])

    videoids = sorted(videoids)
    print(f"Found {len(videoids)} unique videos to process")

    stats = {"ok": 0, "skip": 0, "missing_video": 0, "missing_json": 0, "error": 0}

    for i, videoid in enumerate(videoids):
        result = extract_keyframe(videoid, args.video_dir, args.annotation_dir, args.output_dir)

        if result == "ok":
            stats["ok"] += 1
        elif result == "skip":
            stats["skip"] += 1
        elif result.startswith("missing"):
            stats[result] = stats.get(result, 0) + 1
        else:
            stats["error"] += 1
            if stats["error"] <= 10:
                print(f"  Error on {videoid}: {result}")

        if (i + 1) % 1000 == 0 or (i + 1) == len(videoids):
            print(f"[{i+1}/{len(videoids)}] extracted={stats['ok']} skipped={stats['skip']} "
                  f"missing_video={stats.get('missing_video', 0)} "
                  f"missing_json={stats.get('missing_json', 0)} errors={stats['error']}")

    print(f"\nDone! Extracted {stats['ok']} keyframes, "
          f"skipped {stats['skip']} existing, "
          f"{stats.get('missing_video', 0)} videos not found.")


if __name__ == "__main__":
    main()