Related Experiment Video
Updated: May 2, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
1.8K
DiagSWin: A multi-scale vision transformer with diagonal-shaped windows for object detection and segmentation
1Key Laboratory of Smart Human-Computer Interaction and Wearable Technology of Shaanxi Province, Xidian University, Xi'an, 710071, China.
Summary
Vision Transformers (ViTs) struggle with high computational costs. Diagonal-shaped Window (DiagSWin) attention introduces hybrid scales to efficiently capture multi-scale features, improving performance in computer vision tasks.
Area of Science:
- Computer Vision
- Deep Learning
- Artificial Intelligence
Background:
- Vision Transformers (ViTs) excel at capturing global dependencies via self-attention.
- Global self-attention in ViTs incurs significant computational costs, especially for high-resolution tasks.
- Existing methods for cost reduction often sacrifice long-range modeling capabilities and multi-scale feature extraction.
Purpose of the Study:
- To develop an efficient attention mechanism that addresses the limitations of existing Vision Transformers.
- To introduce a novel approach for capturing multi-scale contextual information within Vision Transformer layers.
- To improve performance on various computer vision tasks while reducing computational complexity.
Main Methods:
- Introduced Diagonal-shaped Window (DiagSWin) attention mechanism for modeling attentions in diagonal regions.
- Implemented hybrid scales within attention layers, allowing tokens to attend to nearby and distant elements at different granularities.
- Developed DiagSWin Transformers, a new variant of Vision Transformer models incorporating the DiagSWin attention.
Main Results:
- DiagSWin Transformers demonstrate superior performance across diverse computer vision tasks.
- A large DiagSWin Transformer achieved 84.4% Top-1 accuracy on ImageNet, outperforming SOTA CSWin Transformer with 40% less computation and model size.
- DiagSWin-Base achieved 51.1 box mAP and 45.8 mask mAP on COCO for object detection/segmentation, and 52.3 mIoU on ADE20K for semantic segmentation.
Conclusions:
- The DiagSWin attention mechanism effectively captures multi-scale context while reducing computational complexity.
- DiagSWin Transformers offer a computationally efficient and high-performing alternative to existing Vision Transformer architectures.
- The proposed method shows significant improvements when used as backbones for object detection, segmentation, and other vision tasks.
