search

Vegeta2020 / SE-SSD

SE-SSD: Self-Ensembling Single-Stage Object Detector From Point Cloud, CVPR 2021.
Apache License 2.0
811 stars 128 forks source link

About student loss #84

Closed daxiongpro closed 2 years ago

daxiongpro commented 2 years ago

Hello, thanks for your code. I find that in your loss:loss = cls_loss_reduced + ious_loss + dir_loss + iou_pred_loss, dir_loss and iou_pred_loss didn't introduced in your paper. dir_loss only mentioned but does not give formular out. And I also find many loss like "cls_pos_loss" are same as your another repo "CIA-SSD", and I want to know is these loss necessary in SE-SSD, because you does not mentioned in SE-SSD

Vegeta2020 commented 2 years ago

@daxiongpro All losses associated with hard targets are necessary to be kept to train the model, as they follow the design of SECOND and "dir_loss" is actually a part of location regresion loss. Besides, "iou_pred_loss" and its related consistency loss can be removed, but they can improve the performance of SE-SSD slightly.