Research map / Architectures and efficiency / Attention and transformers
Advancing Vision Transformer With Enhanced Spatial Priors
Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, Ran He
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 |
|---|---|---|---|
| add Euclidean decay matrix (ablation) | EVT-T · ImageNet-1K | top-1 accuracy: 82.3 -> 83.0 | Improved |
| add Euclidean decay matrix (ablation) | EVT-T · KIT: ADE20K Semantic FPN | mIoU: 46.6 -> 48.3 | Improved |
| use Euclidean instead of Manhattan distance for decay (ablation) | EVT-T · ImageNet-1K | top-1 accuracy: 82.7 -> 83.0 | Improved |
| use Euclidean distance (lower JS divergence to learned attention) (ablation) | EVT-T · ImageNet-1K | JS divergence: 0.20 -> 0.12 | Improved |
| add local context enhancement (LCE) (ablation) | EVT-T · ImageNet-1K | top-1 accuracy: 82.8 -> 83.0 | Improved |
| add conditional position encoding (CPE) (ablation) | EVT-T · ImageNet-1K | top-1 accuracy: 82.9 -> 83.0 | No clear effect |
| add conv stem (ablation) | EVT-T · ADE20K Semantic FPN | mIoU: 47.1 -> 48.3 | Improved |
| swap grouped and dilated attention order (EuSA_g to EuSA_d) (ablation) | EVT-T · ImageNet-1K | top-1 accuracy: 83.0 -> 82.5 | Hurt |
| replace 2D grouping with 1D grouping (same 49 tokens, with EuSA) (ablation) | EVT-S · ImageNet-1K | top-1 accuracy: 83.8 -> 83.9 | No clear effect |
| increase group size with 1D grouping from 49 to 98 tokens (ablation) | EVT-S · ImageNet-1K | top-1 accuracy: 83.9 -> 84.4 | Improved |
| 1D grouping instead of 2D at 512 resolution (direct inference) (ablation) | EVT-T · ImageNet-1K | top-1 accuracy: 81.1 -> 82.8 | Improved |
| use Minkowski p=3 distance instead of Euclidean (ablation) | EVT-Swin-T · COCO Mask R-CNN | box AP: 48.6 -> 44.4 | Hurt |
Closest papers by meaning
- RMT: Retentive Networks Meet Vision Transformers (CVPR 2024 · similarity 0.91)
- RegionViT: Regional-to-Local Attention for Vision Transformers (ICLR 2022 · similarity 0.90)
- MaxViT: Multi-axis Vision Transformer (ECCV 2022 · similarity 0.89)
- LF-ViT: Reducing Spatial Redundancy in Vision Transformer for Efficient Image Recognition (AAAI 2024 · similarity 0.88)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.87)
- QT-ViT: Improving Linear Attention in ViT with Quadratic Taylor Expansion (NeurIPS 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.
