Research map / Architectures and efficiency / Attention and transformers
Making Vision Transformers Efficient From a Token Sparsification View
Shuning Chang, Pichao Wang, Ming Lin, Fan Wang, David Junhao Zhang, Rong Jin, Mike Zheng Shou
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 |
|---|---|---|---|
| Replace image tokens with 16 semantic tokens | DeiT-S · ImageNet-1K val | Top-1 accuracy: 79.8 -> 79.8(+0.0) | No clear effect |
| Replace image tokens with 16 semantic tokens | DeiT-S · ImageNet-1K val | FLOPs: 4.58 -> 1.91(-58%) | Improved |
| Use 9 semantic tokens per window | Swin-S · ImageNet-1K val | Top-1 accuracy: 83.0 -> 83.0(-0.0) | No clear effect |
| Move STGM toward shallower layers (ablation) | Swin-T/S/B · ImageNet-1K val | Top-1 accuracy: 81.3 -> 81.0(-0.3%); 83.0 -> 82.8(-0.2%); 83.5 -> 83.2(-0.3%) | Hurt |
| Use STViT-R recovery and dumbbell units | Swin-S backbone · COCO 2017, Cascade Mask R-CNN 3× | AP^b_s: 35.2 -> 36.7 | Improved |
| Use STViT-R backbone | Swin-S · ADE20K val | mIoU: 49.3 -> 48.3 | Hurt |
| Combine spatial and global initialization (ablation) | STViT-DeiT-S · ImageNet-1K val | Top-1 accuracy: 79.4 -> 79.8 | Improved |
| Replace spatial initialization with global initialization (ablation) | STViT-DeiT-S · ImageNet-1K val | Top-1 accuracy: 79.8 -> 78.7 | Hurt |
| Add learned positional encoding (ablation) | STViT-DeiT-S · ImageNet-1K val | Top-1 accuracy: 79.8 -> 79.7 | Hurt |
| Increase STGM from 2 to 3 transformer layers (ablation) | STViT-DeiT-S · ImageNet-1K val | Top-1 accuracy: 79.8 -> 79.5 | Hurt |
| Add dumbbell units (ablation) | STViT-R-Swin-S · COCO 2017 | AP^b: 51.4 -> 51.8 | Improved |
| Add dumbbell units (ablation) | STViT-R-Swin-S · COCO 2017 | AP^m: 44.4 -> 44.7 | Improved |
Closest papers by meaning
- Tokens-to-Token ViT: Training Vision Transformers From Scratch on ImageNet (ICCV 2021 · similarity 0.92)
- Token Cropr: Faster ViTs for Quite a Few Tasks (CVPR 2025 · similarity 0.91)
- DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification (NeurIPS 2021 · similarity 0.91)
- ImagePiece: Content-aware Re-tokenization for Efficient Image Recognition (AAAI 2025 · similarity 0.91)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · similarity 0.89)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.89)
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.
