Smooth L1

smoothl1loss.pdf

0.13MB

 

 

The Smooth L1 loss is used for doing box regression on some object detection systems, (SSD, Fast/Faster RCNN) according to those papers this loss is less sensitive to outliers, than other regression loss, like L2 which is used on R-CNN and SPPNet. On the Fast RCNN paper, section 2.3 is claimed that the L2 loss need a smaller learning rate to avoid exploding gradients.