MixFormer: Mixing Features across Windows and Dimensions

TL;DR

MixFormer introduces a novel architecture that combines local-window self-attention with depth-wise convolution and bi-directional cross-branch interactions to enhance receptive fields and modeling capabilities in vision tasks.

Contribution

It proposes a new design integrating windowed self-attention with convolution and cross-branch interactions for improved feature mixing in vision models.

Findings

Achieves competitive image classification results with EfficientNet.

Outperforms RegNet and Swin Transformer in accuracy.

Excels in dense prediction tasks with less computational cost.

Abstract

While local-window self-attention performs notably in vision tasks, it suffers from limited receptive field and weak modeling capability issues. This is mainly because it performs self-attention within non-overlapped windows and shares weights on the channel dimension. We propose MixFormer to find a solution. First, we combine local-window self-attention with depth-wise convolution in a parallel design, modeling cross-window connections to enlarge the receptive fields. Second, we propose bi-directional interactions across branches to provide complementary clues in the channel and spatial dimensions. These two designs are integrated to achieve efficient feature mixing among windows and dimensions. Our MixFormer provides competitive results on image classification with EfficientNet and shows better results than RegNet and Swin Transformer. Performance in downstream tasks outperforms its alternatives by significant margins with less computational costs in 5 dense prediction tasks on MS COCO, ADE20k, and LVIS. Code is available at \url{https://github.com/PaddlePaddle/PaddleClas}.