bfshi/DGAM-Weakly-Supervised-Action-Localization

Some questions about THUMOS test features

liming-ai opened this issue · 2 comments

liming-ai commented

hi @bfshi , thanks for your amazing work and congratulates you received many PhD offers.

I noticed that there are some code (e.g. UM, STPN) dropped some test videos due to their false annotation.

I noticed that the provided dataset by you has a txt file named THUMOS14_test_vid_list.txt, I want to know whether the order of feature naming is consistent with the order in this file?

e.g.

video_test_0000004 (the first row in THUMOS14_test_vid_list.txt) -> 1.npy
video_test_0001558 (the last row in THUMOS14_test_vid_list.txt) -> 210.npy

Thanks again and congratulates again, hope you can reply to me!

liming-ai commented

I found that the dataset provided by this repo has removed some falsely annotated videos(i.e. 270, 1496 in test dataset), like UM, STPN did.

zfhxi commented

I also noted this difference, I wonder if it is unfair to compare the experiment results with other methods which don't remove videos.