gulvarol/bsl1k

WLASL matching poses to examples

AmitMY opened this issue · 2 comments

AmitMY commented

I'm looking at the WLASL dataset, which includes 21,083 "instances" of signs.
However, the pose.pkl for that dataset is a list of 20,245 items.

What is the correct way to map between videos and poses?

gulvarol commented

After our download procedure, the videos which were "readable" (according to this line https://github.com/gulvarol/bsl1k/blob/master/misc/wlasl/create_info.py#L111) were saved in the info.pkl which was released here: https://github.com/gulvarol/bsl1k/#supported-datasets
This info.pkl has the metadata for 20,245 videos which corresponds exactly to the poses. We were able to only use these videos in our experiments. Hope this helps.

AmitMY commented

Thanks, I'll try to use info to find the matching videos, and let you know how it goes