gakkiri / augfpn-detectron2 Goto Github PK

Great reimplementation! And I'm wondering how to visualize the spatial attention maps of spatial_attention_conv layer. Have you ever tried it?

Hello, in the original detectron2 architecture, the output of fpn is return {f: res for f, res in zip(self._out_features, results)}, and the output after your modification is return dict(zip(self._out_features, results)), dict(zip(self._out_features, raw_laternals)), then do I still need to modify the features passed in the rpn code? thanks

Hello,repo's author.I wang to add RFA to the one-stage detector,just need to introduce Adaptive pooling? use generated M6 to concat M5?
If possible, I look forward to communicating with you in detail.
Looking forward to your reply!