endernewton/tf-faster-rcnn:候选框是如何跟真实框(ground truth)对应做bbox回归训练的?
1、在RPN网络里通过1*1预测anchors的偏移(rpn_bbox_pred);2、在RPN网络里self._proposal_layer()—>proposal_layer_tf—>bbox_transform_inv_tf(anchors,rpn_bbox_pred)来预测anchors的坐标;3、还是在RPN网络里self._anchor_target_layer()—>anchor_t