Hao Feng

agnostic have 8 bbox_pred num_output(for foreground and background). It may have some difference in mAP. agnostic can be faster and use less memory

@xiaohu2015 I think that background part is used in training. In proposal_target_layer.py, https://github.com/YuwenXiong/py-R-FCN/blob/290dba7fad47f1ec74e2d1675e72cca6efb6cc14/lib/rpn/proposal_target_layer.py#L188 , we use it to get bbox_inside_weights which used in loss_bbox

@xiaohu2015 I think u are right. The bg roi are not used in training bbox. 4(k^2) is enough. However, I try that training with 4 num_output in pascal_voc. It gets...

Hi, sorry, actually I also don't know how to calculate it. But I guess this [project](https://github.com/simbaforrest/vpdetection) may help u.

Sorry, I can't give any comments. It seems abnormal. It's too long ago to remember the details. And I have no free time to reread the paper.