search

dingfengshi / TriDet

[CVPR2023] Code for the paper, TriDet: Temporal Action Detection with Relative Boundary Modeling
MIT License
160 stars 13 forks source link

help with decode_offset function #5

Closed zivnachum closed 1 year ago

zivnachum commented 1 year ago

Hi, thanks for the great code base and paper!

I'm already using Actionformer and now I'm experimenting on TriDet too. When using Actionformer, I made a change in the code so the loss and predictions will be reported in both train and eval mode, and then I can keep track on loss and mAP for train and validation. I tried to make a similar adjustment in TriDet, but encountered an issue in the function decode_offset in meta_archs.py. The fact that there's a different behavior for train and eval makes it harder to get predictions for train too. Any chance you give some explanation on what this function is doing?

Thanks!

dingfengshi commented 1 year ago

Hi, thanks for your attention of our work. This function decode the trident-head output with the formulation (4) (5) (6) (7) in paper. In essence, this function does the same thing during training and validation, but because the input format is different during training and validation (out_offsets is a list containing the features of each FPN level during training. During validation, each FPN feature is calculated separately, so out_offsets is a feature), so adaptation is made in this function.

dingfengshi commented 1 year ago

I have updated some comment for meta_arch.py. Thanks for your reminder!