Why use sigmoid cross entropy instead of softmax in RPN?

[twmht]

Hi,

In the original faster rcnn, you used softmax loss when training rpn. (https://github.com/rbgirshick/py-faster-rcnn/blob/master/models/pascal_voc/VGG16/faster_rcnn_alt_opt/stage1_rpn_train.pt#L447)

in FPN, you use sigmoid cross entropy to measure rpn loss.
(https://github.com/facebookresearch/Detectron/blob/master/lib/modeling/FPN.py#L459)

in my experiment, I found that the RPN recall dropped about 4 point when using sigmoid cross entropy.

So, why using sigmoid cross entropy in FPN? Have you ever tried softmax loss?

thanks!

[rbgirshick]

Using softmax for binary classification is an over parameterization and shouldn't be necessary. When porting from py-faster-rcnn to detectron, I tried both softmax and sigmoid for RPN and obtained similar RPN recall. I have not revisited using softmax for RPN with FPN.

[fw509]

@rbgirshick Is it possible to support Softmax in "rpn_heads.py"?

I tried but didn't succeed. The reason is that when the shape of rpn_cls_logits is (1, 30, H, W) instead of (1, 15, H ,W), the "SpatialNarrowAs" operation cann't be applied to "rpn_labels_int32_wide" since its depth is not the same as rpn_cls_logits.

[Jing-Luo]

I found the same issue here, and it is great to see you put it up. I guess I will try both loss functions to check which one is better.