Formal defintion of G_cls, G_reg, l_cls, l_reg
See original GitHub issueHello @MendelXu!
The paper mentions that G_cls is produced from teacher model by foreground filtering, G_reg is produced from teacher model by box variance filtering; unfortunately the paper doesn’t mention the definition of l_cls
and l_reg
. l_cls(student_candidate_box, teacher_pseudo_boxes)
and l_reg(student_candidate_box, teacher_pseudo_boxes)
still needs to do label assignment. How is this assignment performed? What losses are used?
The paper mentions: Another important benefit of this end-to-end framework is that it allows for greater leverage of the teacher model to guide the training of the student model, rather than just providing “some generated pseudo boxes with hard category labels” as in previous approaches [27, 36]. A soft teacher approach is proposed to implement this insight. In this approach, the teacher model is used to directly assess all the box candidates that are generated by the student model,
rather than providing “pseudo boxes” to assign category labels and regression vectors to these student-generated box candidates
.
It seems that in SoftTeacher, pseudo boxes with hard labels (G_cls, G_reg) are still generated and that some standard IoU-based target-box matching / assignment (l_cls, l_reg) is used. If it’s not the case, could you please bring some clarifications?
It’s possible to recover this information from the code, but some formal definitions could help reading the code as well.
Thank you!
Issue Analytics
- State:
- Created 2 years ago
- Comments:13
Top GitHub Comments
l_cls
andl_reg
is standard loss function for object detection. Concretely, for the detector Faster RCNN,l_cls
is cross-entropy loss. Andl_reg
is L1 loss. And the assignment is still performed in a standard way.Yes.