Research map / Detection and segmentation / CNN features
Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows
Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add shifted window partitioning between consecutive blocks (ablation) | Swin-T · ImageNet-1K | top-1 accuracy: 80.2 -> 81.3 | Improved |
| add shifted window partitioning between consecutive blocks (ablation) | Swin-T · COCO (Cascade Mask R-CNN) | box AP: 47.7 -> 50.5 | Improved |
| add shifted window partitioning between consecutive blocks (ablation) | Swin-T · ADE20K | mIoU: 43.3 -> 46.1 | Improved |
| add relative position bias instead of no position encoding (ablation) | Swin-T · ImageNet-1K | top-1 accuracy: 80.1 -> 81.3 | Improved |
| add relative position bias instead of no position encoding (ablation) | Swin-T · ADE20K | mIoU: 43.8 -> 46.1 | Improved |
| use relative position bias instead of absolute position embedding (ablation) | Swin-T · COCO (Cascade Mask R-CNN) | box AP: 49.0 -> 50.5 | Improved |
| add absolute position embedding on top of relative bias (ablation) | Swin-T · ADE20K | mIoU: 46.1 -> 44.0 | Hurt |
| add absolute position embedding without relative bias (ablation) | Swin-T · ImageNet-1K | top-1 accuracy: 80.1 -> 80.5 | Improved |
| replace sliding window self-attention with shifted windows (ablation) | Swin-T · ImageNet-1K | top-1 accuracy: 81.4 -> 81.3 | No clear effect |
| use cyclic-shift batch computation instead of padding (ablation) | Swin-T · V100 throughput | FPS: 670 -> 755 | Improved |
| use shifted windows instead of sliding windows (kernel) (ablation) | Swin-T · V100 throughput | FPS: 488 -> 755 | Improved |
| replace Performer attention with shifted-window attention (ablation) | Swin-T · ImageNet-1K | top-1 accuracy: 79.0 -> 81.3 | Improved |
Closest papers by meaning
- What Makes for Hierarchical Vision Transformer? (TPAMI 2023 · similarity 0.90)
- Swin Transformer V2: Scaling Up Capacity and Resolution (CVPR 2022 · similarity 0.90)
- HiViT: A Simpler and More Efficient Design of Hierarchical Vision Transformer (ICLR 2023 · similarity 0.89)
- CSWin Transformer: A General Vision Transformer Backbone With Cross-Shaped Windows (CVPR 2022 · similarity 0.88)
- SLAB: Efficient Transformers with Simplified Linear Attention and Progressive Re-parameterized Batch Normalization (ICML 2024 · similarity 0.87)
- Win-Win: Training High-Resolution Vision Transformers from Two Windows (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
