Why there is no ReLU in FPN output

[bowenc0221]

Hi @roytseng-tw ,

Thanks for the nice work. I have a question about the FPN code: why there is no ReLU activation in the Post-hoc scale-specific 3x3 convolutions as well as the lateral connection?

[roytseng-tw]

First of all, there is no ReLU there in Detectron's FPN implementation either. Here's my opinion on why:

1. post-hoc scale-specific 3x3 convs: the outputs are used to extract roi features, so it's better to be dense.
2. lateral convs: the outputs are fused with the features one-level above, so you want them to have same kind of distribution.

[ppwwyyxx]

This exact question is answered in the paper already:

There are no non-linearities in these extra layers, which we have empirically found to have minor impacts

[bowenc0221]

Thanks!