Research map / Architectures and efficiency / Attention and transformers
CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classification
Chun-Fu (Richard) Chen, Quanfu Fan, Rameswar Panda
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 cross-attention fusion instead of no fusion (ablation) | CrossViT-S · ImageNet-1k val (CrossViT-S ablation) | top-1 accuracy (%): 80.2 -> 81.0 | Improved |
| use cross-attention fusion instead of all-attention (ablation) | CrossViT-S · ImageNet-1k val (CrossViT-S ablation) | top-1 accuracy (%): 80.0 -> 81.0 | Improved |
| use cross-attention fusion instead of class token fusion (ablation) | CrossViT-S · ImageNet-1k val (CrossViT-S ablation) | top-1 accuracy (%): 80.3 -> 81.0 | Improved |
| use cross-attention fusion instead of pairwise fusion (ablation) | CrossViT-S · ImageNet-1k val (CrossViT-S ablation) | top-1 accuracy (%): 80.3 -> 81.0 | Improved |
| use CLS tokens for cross-attention instead of averaged patch tokens (ablation) | CrossViT-S · ImageNet-1k val (CrossViT-S ablation) | top-1 accuracy (%): 80.0 -> 81.0 | Improved |
| use patch sizes (12,16) instead of (8,16) (ablation) | CrossViT-S · ImageNet-1k val (CrossViT-S ablation) | top-1 accuracy (%): 80.8 -> 81.0 | Improved |
| widen the small branch to 384 dims (ablation) | CrossViT-S · ImageNet-1k val (CrossViT-S ablation) | top-1 accuracy (%): 81.0 -> 80.1 | Hurt |
| deepen the small branch (ablation) | CrossViT-S · ImageNet-1k val (CrossViT-S ablation) | top-1 accuracy (%): 81.0 -> 80.7 | Hurt |
| stack two cross-attention modules (ablation) | CrossViT-S · ImageNet-1k val (CrossViT-S ablation) | top-1 accuracy (%): 81.0 -> 81.0 | No clear effect |
| use more multi-scale encoders with fewer small-branch encoders (ablation) | CrossViT-S · ImageNet-1k val (CrossViT-S ablation) | top-1 accuracy (%): 81.0 -> 80.9 | Hurt |
| add T2T module to replace linear patch embedding (ablation) | CrossViT-18 · ImageNet-1k val | top-1 accuracy (%): 82.5 -> 83.0 | Improved |
| use convolutional patch embedding (ablation) | CrossViT-9 · ImageNet-1k val | top-1 accuracy (%): 73.9 -> 77.1 | Improved |
Closest papers by meaning
- CrossFormer: A Versatile Vision Transformer Hinging on Cross-scale Attention (ICLR 2022 · similarity 0.92)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · similarity 0.90)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.88)
- ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions (CVPR 2024 · similarity 0.88)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.87)
- A Simple Interpretable Transformer for Fine-Grained Image Classification and Analysis (ICLR 2024 · similarity 0.87)
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.
