Learning-Deep-Learning

SKNet: Selective Kernel Networks

May 2020

tl;dr: A new plug-and-play module of SKNet. Two-stream SENet.

Overall impression

This is a solid work to extend SENet (squeeze and excitation). It chooses adaptive receptive field. Either 3x3 or 5x5.

Compared to inception network, which has multiple parallel path of diff RF, SKNet adaptively chooses which path to focus more.

This inspired ResNeSt and is actually almost exactly the same.

Key ideas

Technical details

Notes