when distilling Faster RCNN with FPN, do we need to firstly assign gt box to different FPN level and then generate imitation region? Or for every FPN level, we generate imitation region according to all gt boxes?

[UcanSee]

[twangnh]

on each level of fpn, only the assigned gt box will be used calculate the imitation region

[UcanSee]

I met with some problems when add your proposed imitation loss to faster rcnn with fpn, I assign gt in this way: one a certain level, for a gt box, If the anchor box which have largest iou with the gt box is not generated from this level and all anchors in this level have no iou with the gt box lager than 0.7, then the gt box should not participate in generating imitation region. if a gt box not satisfy the above two conditions, then the gt box should participate in generating imitation region in paper way. I've got bad results after distillation in above way, did I make something wrong when assigning gt or others?

[ggjy]

I met with some problems when add your proposed imitation loss to faster rcnn with fpn, I assign gt in this way: one a certain level, for a gt box, If the anchor box which have largest iou with the gt box is not generated from this level and all anchors in this level have no iou with the gt box lager than 0.7, then the gt box should not participate in generating imitation region. if a gt box not satisfy the above two conditions, then the gt box should participate in generating imitation region in paper way. I've got bad results after distillation in above way, did I make something wrong when assigning gt or others?

Do you find out how to do the imitation with FPN? I also got bad results in situation such as Res50 based FPN as teacher, Res18 based FPN as student.