Query on THUMOS Inference

[anonyn498]

Hi Author,

Congratulation on the awesome work.

While going through your implementation for THUMOS Dataset, one thing I came to know is you do not use the start and end regression which is claimed to provide the start and end points for the proposal

gtad/gtad_inference.py

Line 36 in 6deb5b1

confidence_map, _, _ = model(input_data)

Is there any reason to do this ?

[frostinassiky]

Hi @anonyn498 . Thanks for the question! It would be interesting to see if the start and end regression work for this problem. According to some recent work (see TCA-Net), action location can be more precise via a second network to regress the boundary.

In G-TAD, I didn't find an evident experimental boost from boundary regression. However, regress the anchor's IoU score is useful.