What to do with corrupted videos

Dear challenge organizers,

there has been some discussion in the forums about seemingly corrupt video data (Link).

We want to ask again how to proceed with these cases. (We are aware that there has been a proposed solution, but we checked it and it turned out to just yield frames from the beginning of the video!!! (See our code here to test it yourselves)). We also checked the video with ffprobe and it contained only 60 frames (ffprobe -show_frames test/vid_1_short.mp4).

Our big question is: In your result json you have generated 62 entries, but we can only produce 60 from the frames available. We can check the number of frames registered in the mp4 container, but that number might be different to the actually available number of frames. What should we do? Please advise !

With kind regards TS

Good question! Also waiting for a clarification ...

Dear Organizers,

any news about my question? We think this is quite an important point to get straight.

With kind regards TS

Hi teams,

We have checked the test set videos from our side and you should not expect corrupted video files. Though we do recommend using the suggested way of reading the videos (using opencv) which should not cause issues regarding inference length. Regardless, we have also put in checks within the evaluation container to not penalize if such small valid differences in gt/predictions occur.

Best, SurgToolLoc 2022 Organizing Committee