Research map / Architectures and efficiency / Attention and transformers
MG-ViT: A Multi-Granularity Method for Compact and Efficient Vision Transformers
Yu Zhang, Yepeng Liu, Duoqian Miao, Qi Zhang, Yiwei Shi, Liang Hu
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 |
|---|---|---|---|
| use global-attention-class (CASR across layers) division instead of random division (ablation) | MG-ViT (DeiT-S) · ImageNet MGIS | top-1 accuracy: 78.9 -> 80.8 | Improved |
| use CASR from last layer only vs lambda-weighted global CASR (ablation) | MG-ViT (DeiT-S) · ImageNet MGIS | top-1 accuracy: 80.5 -> 80.8 | Improved |
| lambda sweep for global class attention (0 to 0.999) (ablation) | MG-ViT (DeiT-S) · ImageNet | top-1 accuracy: 80.3 -> 80.8 | No clear effect |
| add Feature Reuse module (ablation) | MG-ViT (DeiT-S) · ImageNet MGIS | top-1 accuracy: 79.9 -> 80.8 | Improved |
| use zero-padding of tail tokens in feature reuse (ablation) | MG-ViT (DeiT-S) · ImageNet MGIS | top-1 accuracy: 80.7 -> 80.8 | No clear effect |
| match-merge boundary tokens instead of pruning (ablation) | MG-ViT (DeiT-S) · ImageNet | top-1 accuracy: 79.5 -> 80.8 | Improved |
| match-merge instead of K-means token clustering (ablation) | MG-ViT (DeiT-S) · ImageNet | top-1 accuracy: 79.8 -> 80.8 | Improved |
| weighted-average merging instead of keep-one (ablation) | MG-ViT (DeiT-S) · ImageNet | top-1 accuracy: 80.3 -> 80.8 | Improved |
| soft (three-way) slimming instead of hard slimming with merging (ablation) | MG-ViT (DeiT-S) · ImageNet | top-1 accuracy: 80.3 -> 80.8 | Improved |
| merge tokens instead of pruning under soft slimming (ablation) | MG-ViT (DeiT-S) · ImageNet | top-1 accuracy: 80.2 -> 80.8 | Improved |
| multi-granularity splitting on all images (eps 1) vs DeiT-S | DeiT-S · ImageNet | top-1 accuracy: 79.8 -> 81.0 | Improved |
| extend to hierarchical ViT (MG-ViT-A) on PVT-Small | PVT-Small · COCO Mask R-CNN | box AP: 40.4 -> 40.9 | Improved |
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.
- CF-ViT — CF-ViT: A General Coarse-to-Fine Method for Vision Transformer (AAAI 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
- CF-ViT: A General Coarse-to-Fine Method for Vision Transformer (AAAI 2023 · similarity 0.91)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.89)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.88)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.88)
- Outlier-aware Slicing for Post-Training Quantization in Vision Transformer (ICML 2024 · similarity 0.88)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (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.
