Research map / Architectures and efficiency / Attention and transformers
Twins: Revisiting the Design of Spatial Attention in Vision Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace PVT absolute positional encoding with CPE | PVT-Small · ImageNet-1K | Top-1 accuracy (%): 79.8 -> 81.2 | Helped |
| Use spatially separable self-attention with local and global attention | Twins-SVT-Small · ImageNet-1K | Top-1 accuracy (%): 81.7 | Helped |
| Use conditional positional encoding in PVT | PVT-Small · ADE20K val, Semantic FPN 80k | mIoU (%): 39.8 -> 44.3 | Helped |
| Use spatially separable self-attention | Twins-SVT-Large · ADE20K val, UperNet 160k MS | mIoU (%): 49.7 -> 50.2 | Helped |
| Use conditional positional encoding in PVT | PVT-Small · COCO val2017, RetinaNet 1x | AP: 40.4 -> 43.0 | Helped |
| Use only locally-grouped attention, omitting global attention (ablation) | Twins-SVT-Small · ImageNet-1K | Top-1 accuracy (%): 81.7 -> 76.9 | Hurt |
| Add a global attention layer at the final stage (ablation) | LSA-only configuration · ImageNet-1K | Top-1 accuracy (%): 76.9 -> 80.5 | Helped |
| Use LLG blocks in the middle stages (ablation) | Twins-SVT-Small · ImageNet-1K | Top-1 accuracy (%): 81.5 | No clear effect |
| Replace regular strided convolution with average pooling for GSA subsampling (ablation) | Twins-SVT-Small · ImageNet-1K | Top-1 accuracy (%): 81.7 -> 81.2 | Hurt |
| Replace regular strided convolution with separable convolution for GSA subsampling (ablation) | Twins-SVT-Small · ImageNet-1K | Top-1 accuracy (%): 81.7 -> 81.2 | Hurt |
| Replace Swin-T relative positional encoding with CPE (ablation) | Swin-T · ImageNet-1K | Top-1 accuracy (%): 81.3 -> 81.2 | Hurt |
| Replace Swin-T relative positional encoding with CPE (ablation) | Swin-T · COCO val2017, RetinaNet | AP: 41.5 -> 41.3 | Hurt |
Closest papers by meaning
- ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases (ICML 2021 · similarity 0.89)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.89)
- Neighborhood Attention Transformer (CVPR 2023 · similarity 0.88)
- FasterViT: Fast Vision Transformers with Hierarchical Attention (ICLR 2024 · similarity 0.88)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.87)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.87)
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.