Open escorciav opened 5 years ago
something like this should fix it
for clip_ind in range(num_clips_i):
if clip_ind == (num_clips_i - 1):
clip_s = dataset._video_duration(video_id)
else:
clip_s = clip_length * (clip_ind + 0.5)
clip_ge_start = clip_s >= proposals_[:, 0]
clip_le_end = clip_s <= proposals_[:, 1]
Next: Launch evaluation for ApproxCAL again
Please provide me with the final version of the solution to be changed in CVPR2020 branch.
The fact that we have clips that don't span any proposal is related to the way we map both here:
https://github.com/escorciav/moments-retrieval/blob/69e4dc05f87e740b834e350061d90c55b070fce8/corpus.py#L104-L113
In a nutshell, we aren't handling the right boundary precisely.
If the center of the last clip is greater than
duration - 0.5 * clip_length
, it won't span any moment.