About the focal loss layer

[suzhenghang]

Hi @unsky , The performance in your experiment is amazing. By the way, did you only replace the SoftmaxWithLoss with the focal loss layer in RPN layer or in both RPN and Fast RCNN?

[unsky]

just fast rcnn

[suzhenghang]

@unsky Thanks, did you try to replace the SoftmaxWithLoss in RPN with the focal loss (2 class, fore/back ground)? As the focal loss is used to process the unbalanced foreground-backgroind class.

[twmht]

@unsky and @suzhenghang

I have tried focal loss to rpn, the loss converges very quickly. but the rpn's recall is very low.

I think the problem is that the loss is very low at the beginning (the loss drops 10e-3 within 2k iterations).

thus cause the gradient is very low when doing backward propagation.

here is the loss historgram

Does anyone who applies focal loss to rpn network and works well?