faster rcnn 损失函数

损失函数分为框预测回归和分类loss，每个都分为rpn和rcnn。

rpn损失：

_anchor_target_layer 给图像内部的anchors分配rpn_labels(1或0，前景或背景)，再和rpn_cls_score做交叉熵，此时只关注物体是前景/背景，不管物体的类别

rpn_cross_entropy = tf.reduce_mean(
        tf.nn.sparse_softmax_cross_entropy_with_logits(logits=rpn_cls_score, labels=rpn_label))

对图像内部的框做回归 

# 计算anchor平移和放缩的幅度target，和所属的类比较
  bbox_targets = _compute_targets(anchors, gt_boxes[argmax_overlaps, :])
......
# bbox_targets
  bbox_targets = bbox_targets \
    .reshape((1, height, width, A * 4))

  rpn_bbox_targets = bbox_targets

rpn_loss_box = self._smooth_l1_loss(rpn_bbox_pred, rpn_bbox_targets, rpn_bbox_inside_weights,
                                          rpn_bbox_outside_weights, sigma=sigma_rpn, dim=[1, 2, 3])

rcnn损失：

_proposal_target_layer 选出32个前景和96个背景，前景_crop_pool_layer之后做分类(cls_score)和回归(bbox_pred)

pool5 = self._crop_pool_layer(net_conv, rois, "pool5")

      # RCNN, class loss
      cross_entropy = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=cls_score, labels=label))

      # RCNN, bbox loss
      loss_box = self._smooth_l1_loss(bbox_pred, bbox_targets, bbox_inside_weights, bbox_outside_weights)

别人的终究是别人的，自己懂了才是自己的。（看了那么多博客，好像没什么帮助，还得靠自己悟）