Research map / Architectures and efficiency / Attention and transformers
Saliency-Driven Token Merging for Vision Transformers
Weiying Xie, Xiaoyu Chen, Xin Zhang, Chenhe Hao, Jitao Ma, Yunsong Li, Leyuan Fang
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 |
|---|---|---|---|
| combine saliency outlier score with class attention (alpha=beta=0.5) instead of saliency alone (ablation) | SAD-TM on DeiT-Small · ImageNet-1k, 2.6 GFLOPs | Top-1 accuracy: 78.02 -> 79.27 | Improved |
| combine class attention with saliency instead of attention alone (ablation) | SAD-TM on DeiT-Base · ImageNet-1k, 10.1 GFLOPs | Top-1 accuracy: 80.48 -> 80.60 | Improved |
| combine class attention with saliency instead of attention alone (ablation) | SAD-TM on DeiT-Tiny · ImageNet-1k, 0.8 GFLOPs | Top-1 accuracy: 71.47 -> 71.58 | Improved |
| add deferred merging (SAD-TM-DM) | SAD-TM on DeiT-Small · ImageNet-1k, 3.0 GFLOPs | Top-1 accuracy: 79.52 -> 79.71 | Improved |
| add deferred merging (SAD-TM-DM) | SAD-TM on MAE-Huge · ImageNet-1k, ~41% FLOPs reduction | Top-1 accuracy: 86.43 -> 86.58 | Improved |
| add deferred merging (SAD-TM-DM) | SAD-TM on LV-ViT-S · ImageNet-1k, 3.9 GFLOPs | Top-1 accuracy: 82.45 -> 82.60 | Improved |
| use saliency outlier scores computed by a different DeiT model (ablation) | SAD-TM on DeiT-Tiny/Small/Base · ImageNet-1k | Top-1 accuracy: at most 0.12% difference | No clear effect |
| replace Pearson correlation with cosine similarity (ablation) | SAD-TM on DeiT-Base · ImageNet-1k, 10.1 GFLOPs | Top-1 accuracy: 80.60 -> 80.44 | Hurt |
| replace Pearson correlation with Euclidean distance or Mahalanobis distance (ablation) | SAD-TM on DeiT-Small · ImageNet-1k, 2.5 GFLOPs | Top-1 accuracy: 79.15 -> 79.10 / 79.11 | No clear effect |
| replace Pearson correlation with Mahalanobis distance (ablation) | SAD-TM on DeiT-Tiny · ImageNet-1k, 0.8 GFLOPs | Top-1 accuracy: 71.58 -> 71.61 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiffRate — DiffRate : Differentiable Compression Rate for Efficient Vision Transformers (ICCV 2023) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning to Merge Tokens via Decoupled Embedding for Efficient Vision Transformers (NeurIPS 2024 · similarity 0.92)
- Video Token Merging for Long Video Understanding (NeurIPS 2024 · similarity 0.90)
- vid-TLDR: Training Free Token Merging for Light-weight Video Transformer (CVPR 2024 · similarity 0.89)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · similarity 0.89)
- Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement (CVPR 2025 · similarity 0.88)
- Token Transformation Matters: Towards Faithful Post-hoc Explanation for Vision Transformer (CVPR 2024 · similarity 0.88)
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.
