Loss function and positive and negative sample allocation in target detection: retinanet and focal loss

RetinaNet

In the field of target detection , For the first time, the accuracy of single-stage algorithm exceeds that of two-stage algorithm , Namely RetinaNet.

Network structure ：

The network structure designed by the author is not very innovative , This is what the article says ：

The design of our RetinaNet detector shares many similarities with
previous dense detectors, in particular the concept of ‘anchors’
introduced by RPN [3] and use of features pyramids as in SSD [9] and
FPN [4]. We emphasize that our simple detector achieves top results
not based on innovations in network design but due to our novel loss.

Detection head is Classification and BBox Regression decoupling Of , And it is based on anchor frame , after FPN After that, five characteristic maps with different scales are output , Each layer corresponds to 32~512 Anchor frame of scale , And each layer is based on scale and ratios There are different combinations of 9 Seed anchor frame , Finally, the anchor frame size of the whole network is 32 ~ 813 Between . Use the offset predicted by the network relative to the anchor box to calculate BBox Methods and Faster R-CNN identical . The following figure is the picture of thunderbolt .

The structure diagram in the paper is as follows , Only by FPN The characteristic diagrams of three scales are drawn ,W,H,K,A Respectively represent the width of the feature map , high , Number of categories （ Does not contain background classes ）, Number of anchor frames （9）.

Positive and negative samples match

Positive sample ： Predicted BBox And gt IoU>=0.5,
Negative sample ： Predicted BBox And gt IoU<0.4,
Other samples are discarded

prospects , Background quantity imbalance

CE Variants of loss

The biggest innovation of this work ：Focal loss, Rewrite the classic cross entropy loss , Apply to class subnet Branch , The weight of the loss of easily classified samples is greatly reduced , Beautiful form . In the paper $\gamma$ Recommended 2, if $\gamma$ take 0, be FL It degenerates into CE.

Loss ：

The first category loss is to calculate all samples （ Including positive and negative ） Of Focal loss, Then remove the number of positive samples $N_{pos}$.BBox The return loss is Fast R-CNN Proposed in smooth L1 loss.

To be continued