Why did solo use sigmoid focal loss not softmax_cross_entropy loss?

WXinlong / SOLO

SOLO and SOLOv2 for instance segmentation, ECCV 2020 & NeurIPS 2020.

Other

1.69k stars 307 forks source link

Closed Sunting78 closed 3 years ago

Sunting78 commented 4 years ago

I think one grid is only one cate label. Why not use softmax_cross_entropy loss ？

WXinlong commented 4 years ago

@Sunting78 Please refer to RetinaNet.