Question about frame ordering in the agent disappearance metric

[phymhan]

Hi, thanks for releasing DrivingGen!

I noticed that video_a_missing.py loads frames with:

gen_video_path = [
    os.path.join(img_dirs[scene_id], f)
    for f in os.listdir(img_dirs[scene_id])
]
sorted(gen_video_path)

Since sorted() returns a new list, gen_video_path remains in filesystem order. Later, temporal frame IDs are used to index this list, for
example:

first_img = cv2.imread(gen_video_path[first_fid])

On our filesystem, os.listdir() is not chronological, so the frame ID can point to a different image.

Would assigning the sorted result, or directly constructing paths from frame IDs, better match the intended behavior?

gen_video_path = [
    os.path.join(img_dirs[scene_id], f"{i:05d}.png")
    for i in range(101)
]

Thanks!

[youngzhou1999]

Hi, thanks for your detailed assessment of our codebase!

It's definitely better to do that.