Potential data mismatch and incorrect annotations

[Achtoria]

Thanks for your great work on this dataset.
However, I noticed some potential issues with the data. For example, regarding Ego4D_40a7470c-f3ed-41d4-956a-4b65f97e895a_ep_000100, the action and language annotations for the entire video appear to be incorrect.
Additionally, some episodes seem to exceed the frame range of the video file. I suspect that the annotations might be mapped to the wrong video.
Here are the results using the official visualization tool:

Ego4D_40a7470c-f3ed-41d4-956a-4b65f97e895a_ep_000100.mp4

[arnoldland]

Hi, thanks for your interest! Could you clarify how you processed the videos? Are there any operations that might cause changes in the number of frames? In my case, the rendered undistorted videos look as follows, and I do not observe any misalignment or cases where an episode exceeds the length of the video file.

Ego4D_40a7470c-f3ed-41d4-956a-4b65f97e895a_ep_000100.mp4

In addition, I also attach the directly downloaded Ego4D video (without undistortion) for comparison:

Ego4D_40a7470c-f3ed-41d4-956a-4b65f97e895a_ep_000100_before_undistortion.mp4

[Achtoria]

Thanks for your fast response! I find when running the undistortion script with multiple workers, the concatenate_ts_files function uses a hardcoded filename 'inputs.txt' to store the list of TS chunks for FFmpeg. This creates a race condition so the output videos have correct filenames but incorrect/swapped content (i.e., Video_A.mp4 contains frames from Video_B).

def concatenate_ts_files(folder, video_name, batch_counts):
    """Concatenate batch TS files into final mp4."""
    inputs_path = os.path.join(folder, 'inputs.txt')

    # Create a file list for ffmpeg
    with open(inputs_path, 'w') as f:
        for i in range(batch_counts):
            f.write(f"file '{video_name}_b{i:04d}.ts'\n")

    # Merge using ffmpeg concat demuxer
    ffmpeg.input(inputs_path, format='concat', safe=0).output(
        os.path.join(folder, f'{video_name}.mp4'),
        c='copy'
    ).run()

    # Cleanup temporary TS files and list file
    for i in range(batch_counts):
        os.remove(os.path.join(folder, f"{video_name}_b{i:04d}.ts"))
    os.remove(inputs_path)

[arnoldland]

Thank you for raising this issue! Our example code currently only supports single-process execution, and enabling multi-process support would require some modifications (though not extensive ones). To improve usability and ease of modification, I will update the concatenate_ts_files part of the code to avoid the video corruption issue you mentioned.