@@ -98,7 +98,7 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker {
98
98
" This is a 4-D tensor with shape of [N, C, H, W]."
99
99
" H and W should be same, and the second dimention(C) stores"
100
100
" box locations, confidence score and classification one-hot"
101
- " key of each anchor box" );
101
+ " keys of each anchor box" );
102
102
AddInput (" GTBox" ,
103
103
" The input tensor of ground truth boxes, "
104
104
" This is a 3-D tensor with shape of [N, max_box_num, 5], "
@@ -179,6 +179,11 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker {
179
179
box coordinates (w, h), and sigmoid cross entropy loss is used for box
180
180
coordinates (x, y), confidence score loss and classification loss.
181
181
182
+ Each groud truth box find a best matching anchor box in all anchors,
183
+ prediction of this anchor box will incur all three parts of losses, and
184
+ prediction of anchor boxes with no GT box matched will only incur objectness
185
+ loss.
186
+
182
187
In order to trade off box coordinate losses between big boxes and small
183
188
boxes, box coordinate losses will be mutiplied by scale weight, which is
184
189
calculated as follow.
0 commit comments