Research map / Architectures and efficiency / Attention and transformers
CF-ViT: A General Coarse-to-Fine Method for Vision Transformer
Mengzhao Chen, Mingbao Lin, Ke Li, Yunhang Shen, Yongjian Wu, Fei Chao, Rongrong Ji
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 coarse-to-fine inference with η=0.5 | DeiT-S · ImageNet val | FLOPs (G): 4.6 -> 1.8 | Improved |
| Use coarse-to-fine inference with η=0.63 | LV-ViT-S · ImageNet val | FLOPs (G): 6.6 -> 3.1 | Improved |
| Use coarse-to-fine inference with η=0.5 | DeiT-S · ImageNet val | Top-1 accuracy (%): 79.8 -> 79.8 | No clear effect |
| Use global class attention instead of last-layer attention for region selection (ablation) | CF-ViT · ImageNet, fine inference | Top-1 accuracy (%): 80.3 -> 80.8 | Improved |
| Use informative global class-attention regions rather than random regions (ablation) | CF-ViT · ImageNet, fine inference | Top-1 accuracy (%): 79.6 -> 80.8 | Improved |
| Select high-attention rather than low-attention regions for fine splitting (ablation) | CF-ViT · ImageNet, fine inference | Top-1 accuracy (%): 77.6 -> 80.8 | Improved |
| Add feature reuse to the fine-stage tokens (ablation) | CF-ViT · ImageNet, fine inference | Top-1 accuracy (%): 80.0 -> 80.8 | Improved |
| Exclude the [class] token from feature reuse (ablation) | CF-ViT · ImageNet, fine inference | Top-1 accuracy (%): 80.2 -> 80.8 | Improved |
| Exclude uninformative tokens from feature reuse (ablation) | CF-ViT · ImageNet, fine inference | Top-1 accuracy (%): 80.6 -> 80.8 | Improved |
| Use an MLP rather than a linear layer for feature reuse (ablation) | CF-ViT · ImageNet, fine inference | Top-1 accuracy (%): 80.6 -> 80.8 | Improved |
| Increase the informative-patch rate α (ablation) | CF-ViT · ImageNet, fine inference | Top-1 accuracy (%): 80.4 -> 81.4 | Improved |
| Increase the informative-patch rate α (ablation) | CF-ViT · ImageNet, fine inference | FLOPs (G): 3.7 -> 5.5 | Hurt |
Later papers that compared with this method
2 later accepted papers report CF-ViT as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- LF-ViT: Reducing Spatial Redundancy in Vision Transformer for Efficient Image Recognition (AAAI 2024 · better on 4 of 4 shared results)
- MG-ViT: A Multi-Granularity Method for Compact and Efficient Vision Transformers (NeurIPS 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
- MG-ViT: A Multi-Granularity Method for Compact and Efficient Vision Transformers (NeurIPS 2023 · similarity 0.91)
- CvT: Introducing Convolutions to Vision Transformers (ICCV 2021 · similarity 0.90)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.90)
- Your ViT is Secretly an Image Segmentation Model (CVPR 2025 · similarity 0.88)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.88)
- LF-ViT: Reducing Spatial Redundancy in Vision Transformer for Efficient Image Recognition (AAAI 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.
