tianzhi0549 / FCOS

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)
https://arxiv.org/abs/1904.01355
Other
3.28k stars 630 forks source link

bbox_reg.append(bbox_pred * self.fpn_strides[l]) #211

Closed jszgz closed 4 years ago

jszgz commented 4 years ago

这一句是什么意思呢,为什么bbox_reg只有测试的时候需要乘以stride, 不太明白是什么意思, 另外,看这里的self.norm_reg_targets应该是True的, 那作者就是没有用论文里说的下面else里面的exp()函数吧, 还有就是论文里作者说回归的都是正数, 所以直接用exp, 可是看这个if 的分支里面是先relu了,然后也没用exp, 不太懂是什么逻辑呢 https://github.com/tianzhi0549/FCOS/blob/master/fcos_core/modeling/rpn/fcos/fcos.py#L112

tianzhi0549 commented 4 years ago

@jszgz when training, we instead normalize the training targets here. https://github.com/tianzhi0549/FCOS/blob/dd4ca761f439a48a8c3003d1be74e9f85c050534/fcos_core/modeling/rpn/fcos/loss.py#L142

jszgz commented 4 years ago

Thx