MixFormer: Mixing Features across Windows and Dimensions
This work addresses performance bottlenecks in vision transformers for researchers and practitioners, offering an incremental improvement over existing methods like Swin Transformer.
The paper tackles the limited receptive field and weak modeling capability of local-window self-attention in vision tasks by proposing MixFormer, which combines local-window self-attention with depth-wise convolution and bi-directional interactions across branches, achieving competitive results in image classification and outperforming alternatives in 5 dense prediction tasks with less computational cost.
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}.