Datasets:
Problem with the annotations in the FitAQA dataset
While going through the dataset, I noticed that in the multiple_choice subset the video_name field does not match the descriptions and labels provided for each of the dataset instances.
As an example, the first case states that video 00015094.mp4 shows a posterior deltoid stretch, but when I checked that video in the QEVD-FIT-300k dataset it bears no relation to that: it shows a person lying on their back (which is also what the dataset's original annotations indicate).
Thanks for pointing this out. It may be caused by differences between QEVD release versions, where the same filename maps to different videos.
We’ve added a video_sha256 field for the exact video files used in FitAQA, which may help you remap the videos to your QEVD release by matching checksums.