I’ve been trying to replicate the results from ViViT in which you are required to sample 32 frames from videos in the Kinetics dataset. However I have been struggling with a weird issue. Basically, the VideoClip class returns one extra frame that what I require. A similar issue was referenced in issue #1884, but there it returned less than expected.
I have checked and the videos don’t seem to be corrupted. Also it seems to happen to every video, thought I haven’t checked all of them (there’s quite a few).
I know this is probably an issue for Github and I have commented there, but it seems like the video api is not as widely maintained as other parts of the ecosystem and I figured people here are more active.