search

Megvii-BaseDetection / DeFCN

End-to-End Object Detection with Fully Convolutional Network
Apache License 2.0
495 stars 36 forks source link

Issue about 3DMF #10

Closed Hwang64 closed 3 years ago

Hwang64 commented 3 years ago

The filtered feature after 3DMF is added by the original one to form the final output:

https://github.com/Megvii-BaseDetection/DeFCN/blob/a82393e290455fd11a1a088723a8050791b44c15/playground/detection/coco/poto.res50.fpn.coco.800size.3x_ms.3dmf/fcos.py#L607

It is a little contradictory to the purpose of the 3DMF which is used to filter the "unnecessary" point in the feature. If the original feature is used to form the final output, the output will not be a sparse feature map, and it may bring some adverse effects to the results without NMS processing.

wjfwzzc commented 3 years ago

The experiments and visualization show that the feature maps become more sparse. Even so, our solution is of course not the perfect one. It's always welcomed to bring up new improvements.