ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks

January 2021

tl;dr: Train with 3x3, 3x1 and 1x3, but deploy with fused 3x3.

Overall impression

This paper take the idea of BN fusion during inference to a new level, by fusing conv kernels. It has no additional hyperparameters during training, and no additional parameters during inference, thanks to the fact that additivity holds for convolution.

It directly inspired RepVGG, a follow-up work by the same authors.

Key ideas


Technical details