Closed dagongji10 closed 4 years ago
Though I want to unified the conf_loss
as the same parameter setting as other losses, I found the performance would be better by not involving main_conf_loss
into conf_loss
in my experiments. I cannot explain this phenomenon in theory. Maybe you can try in your case.
Thanks.
Hi @HYPJUDY ,
also I noticed that you have used Smooth-L1 loss for the Overlap-loss, but in the paper you mentioned that you used Mean-Squared-Error loss for Overlap-loss. Can you please explain why the difference exists?
Hi @atique81 , sorry for the inconsistency. The code of smooth-L1 loss is directly adopted from SSAD and we didn't clarify this detail in paper. "Smooth L1-loss can be interpreted as a combination of L1-loss and L2-loss"ref. Smooth-L1 loss may perform better than MSE (I didn't do ablation study), but their motivation is similar: to minimize the gap between predicted values and ground truth values. So it should be fine to use both.
Hello @HYPJUDY , Thanks for the clarification.
In your paper, I see overlap loss is , it conbine main_branch and proposal_branch. But in the code it is , it only use proposal_branch. Why ? Is there any tips ?