Closed dongzhenguo2016 closed 4 years ago
Both is fine. The only difference is class aware and class agnostic proposal.
The parameters in RpnParam.bbox_target are used for generating regression bbox targets in rcnn stage. And there are class aware and class agnostic setting in rcnn stage.
在cascade_r101v1_fpn_1x.py中 class RpnParam: class bbox_target: num_reg_class = 2 # RPN只判断proposal是目标还是背景。 但是在faster_r101v1_fpn_1x.py中 class RpnParam: class bbox_target: num_reg_class = 81 # 为什么是81呢?我觉得应该是2才行呀?为何是coco数据集的类别数+1? thanks everyone! happy new year!