Open tokisakikaguya opened 9 months ago
Hi Tokisaki,
Please consider referring to issues #1 and #6 of the action recognition repo.
TL;DR: We use the 2D backbone of a pretrained TSM (trained on video with shifting operations) for extracting the per-frame features. We repurpose the RU-LSTM feature extraction script, by replacing TSN with TSM and using TSM image transforms.
Please let us know if you face any issues.
Best, Dibyadip
Hi! Thank you very much for your reply! I have learned the solution to the problem through your reply! Best, Tokisak
Hello, I noticed that you wrote "TempAgg requires per-frame features as input. TSM (8-frame input) has been used for extracting 2048-D per-frame features" in the Feature section. I would like to know in what form the 8 frames are input during the actual feature extraction process, as well as the more specific single-frame feature extraction strategy. Can you please give me an answer? Thanks!