Research map / Detection and segmentation / CNN features
VSA: Learning Varied-Size Window Attention in Vision Transformers
Qiming Zhang, Yufei Xu, Jing Zhang, Dacheng Tao
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 VSA in place of fixed window attention | Swin-T · ImageNet-1k val (224) | top-1 accuracy: 81.2 -> 82.3 | Improved |
| add VSA to the model at 384 input | Swin-T · ImageNet-1k val (384) | top-1 accuracy: 81.4 -> 83.2 | Improved |
| add VSA to the model at 480 input | Swin-T · ImageNet-1k val (480) | top-1 accuracy: 81.5 -> 83.4 | Improved |
| add VSA to ViTAEv2-S | ViTAEv2-S · ImageNet-1k val (224) | top-1 accuracy: 82.2 -> 82.7 | Improved |
| add VSA to Mask RCNN 1x backbone | Swin-T / ViTAEv2-S · COCO val Mask RCNN 1x | box AP: 43.7 -> 45.6 | Improved |
| add VSA to Mask RCNN 3x backbone | Swin-T / ViTAEv2-S · COCO val Mask RCNN 3x | box AP: 46.0 -> 47.5 | Improved |
| add VSA to Cascade RCNN backbone | Swin-T · COCO val Cascade RCNN | box AP: 50.2 -> 51.4 | Improved |
| add VSA to UperNet backbone | Swin-T · Cityscapes 512x1024 40k | mIoU: 78.9 -> 80.8 | Improved |
| add VSA at stage 4 only (ablation) | Swin-T · ImageNet-1k val (ablation) | top-1 accuracy: 81.2 -> 81.4 | Improved |
| extend VSA from stage 4 to stages 3-4 (ablation) | Swin-T · ImageNet-1k val (ablation) | top-1 accuracy: 81.4 -> 81.9 | Improved |
| extend VSA to all four stages (ablation) | Swin-T · ImageNet-1k val (ablation) | top-1 accuracy: 82.1 -> 82.3 | Improved |
| add VSR (varied-size window regression) alone (ablation) | Swin-T · ImageNet-1k val (ablation) | top-1 accuracy: 81.2 -> 81.6 | Improved |
Closest papers by meaning
- Multi-Scale Representations by Varying Window Attention for Semantic Segmentation (ICLR 2024 · similarity 0.86)
- LookHere: Vision Transformers with Directed Attention Generalize and Extrapolate (NeurIPS 2024 · similarity 0.85)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.85)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.85)
- Towards Long-window Anchoring in Vision-Language Model Distillation (AAAI 2026 · similarity 0.85)
- A-VL: Adaptive Attention for Large Vision-Language Models (AAAI 2025 · similarity 0.85)
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.
