Some Questions about the VLR ?

[ttjjmm]

1. I notice that you separate the VLR and MDR in an ATSS manner, but if I use some other label assignments like OTA or TOOD, should I split the valuable distillation region via quality metrics to keep in step with different LA methods? (set a quality metric threshold)
2. Region Weighting in your paper is from the Student regression feature map or Teacher or the combination of the two?
3. Great idea and extend the probabilistic attribute of DFL to distillation, fantastic !!!

[HikariTJU]

1.分开只是为了方便写代码而已，理论上应该没区别吧。 2.好想法，没试过用teacher的weight，还是用的student

[Zzh-tju]

1. If you want to combine the VLR with some dynamic label assignments, like OTA, AutoAssign, TOOD, you can 1) sort the quality metric and choose 1--top k1 to be positives and top k1--top k2 to be VLRs; or 2) calculate another quality metric and give the localization score a wider range. All you have to ensure is the VLR focuses more on localization and LD has a wider distillation region than KD.
2. Region Weighting is from student's feature map, because you train the student detector, right?

[Tongfengyu]

作者你好，我想请问一下有价值定位区域选择算法那里，生成的是掩码（01）还是权重呢？ 我看你们论文里面是掩码，然后代码里面是权重（Diou）。 谢谢你的回复

[HikariTJU]

掩码， 最后算loss乘以权重