# Swin Transformer: Hierarchical Vision Transformer using Shifted Windows

**Authors:** Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo  
**arXiv:** [2103.14030](https://arxiv.org/abs/2103.14030)  
**Date:** 2021-03-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2103.14030v2/

## Key points

- The Swin Transformer is a novel vision Transformer designed for computer vision tasks.
- It utilizes a hierarchical architecture with Shifted Windows for efficient computation.
- The innovative windowing scheme restricts self-attention calculations to non-overlapping local windows while enabling cross-window connections.
- The Swin Transformer achieves impressive results in image classification, object detection, and semantic segmentation tasks.
- Outperforms previous state-of-the-art models by significant margins in various vision tasks.
- Offers a powerful solution for computer vision tasks with its versatility and efficiency using shifted windows.
- Represents a significant advancement in leveraging Transformer principles for visual data processing.

## Summary

, , , , 

The Swin Transformer: A Versatile Backbone for Computer Vision Tasks

The Swin Transformer, a novel vision Transformer introduced in this paper, offers a versatile backbone for computer vision tasks. Unlike traditional Transformers designed for language processing, adapting them to the visual domain presents unique challenges due to differences in scale and resolution between images and text. To address these challenges, the authors propose a hierarchical Transformer architecture that utilizes Shifted Windows for efficient computation. This innovative windowing scheme restricts self-attention calculations to non-overlapping local windows while enabling cross-window connections. The hierarchical design of the Swin Transformer allows for modeling at various scales with linear computational complexity relative to image size. This flexibility makes it suitable for a wide range of vision tasks, including achieving impressive results in image classification (87.3 top-1 accuracy on ImageNet-1K) and dense prediction tasks such as object detection (58.7 box AP and 51.1 mask AP on COCO test-dev) and semantic segmentation (53.5 mIoU on ADE20K val). Notably, the Swin Transformer outperforms previous state-of-the-art models by significant margins (+2.7 box AP and +2.6 mask AP on COCO, and +3.2 mIoU on ADE20K), showcasing the potential of Transformer-based architectures as effective vision backbones. Moreover, the shifted window approach proves beneficial not only for vision tasks but also for all-MLP architectures in general. <break>

In summary, the Swin Transformer offers a powerful solution for computer vision tasks with its hierarchical architecture and efficient computation using shifted windows. Its success across various vision tasks highlights its potential as an effective vision backbone. The availability of code and models publicly accessible further enhances its utility, making it a valuable framework developed by Microsoft researchers Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. <break>

In contrast to advancements in network architectures in natural language processing (NLP), where Transformers have become prevalent due to their ability to model long-range dependencies using attention mechanisms, the Swin Transformer represents a significant step forward in leveraging similar principles for visual data processing with remarkable success across various vision tasks.