Swing V2: towards a larger model with larger capacity and higher resolution

        Swin Transformer In target detection 、 Examples segmentation and other computer vision tasks have achieved SOTA Performance of . Swin Transformer The field of computer vision has been broken CNN（ Convolutional neural networks ） long-term “ rule ” The situation of , It accelerates the transformation of the basic model architecture in the field of computer vision , This work is therefore To obtain the 2021 year ICCV Best Paper Award —— Mar prize .

        Swin Transformer Two key concepts are introduced to solve the original problem ViT Problems faced —— Hierarchical feature mapping and window attention conversion .

Network structure ( Innovation points )

• Swin Transformer The hierarchical construction method similar to that in convolutional neural network is used （Hierarchical feature maps）, For example, there is down sampling of the image in the feature map size 4 Times ,8 Times and 16 Times , In this way backbone It is helpful to build a target detection system on this basis , Instance segmentation and other tasks . And before Vision Transformer It is directly down sampling from the beginning 16 times , The following characteristic diagram also keeps the sampling rate unchanged .
• stay Swin Transformer Used in Windows Multi-Head Self-Attention(W-MSA) The concept of , For example, in the figure below 4 Double down sampling and 8 Times down sampling , The feature graph is divided into multiple disjoint regions （Window）, also Multi-Head Self-Attention Only in each window （Window） Inside . be relative to Vision Transformer Directly to the whole （Global） The characteristic diagram is used for Multi-Head Self-Attention, The purpose of this is to reduce the amount of calculation , Especially when the shallow characteristic map is very large . Although this reduces the amount of calculation, it will also isolate the information transmission between different windows , So in the paper, the author puts forward Shifted Windows Multi-Head Self-Attent