Research map / Architectures and efficiency / Attention and transformers
ResT: An Efficient Transformer for Visual Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| ResT backbone vs PVT at similar complexity | vision Transformer backbone · ImageNet-1k | top-1 accuracy: 75.1 -> 79.6 | Helped |
| ResT vs Swin at similar complexity | vision Transformer backbone · ImageNet-1k | top-1 accuracy: 81.3 -> 81.6 | Helped |
| ResT-Small vs PVT-T backbone | RetinaNet · COCO val2017 | box AP: 36.7 -> 40.3 | Helped |
| use ResT stem (three 3x3 convs) instead of PVT stem (ablation) | ResT-Lite · ImageNet-1k | top-1 accuracy: 71.96 -> 72.88 | Helped |
| use ResT stem instead of ResNet stem (ablation) | ResT-Lite · ImageNet-1k | top-1 accuracy: 72.24 -> 72.88 | Helped |
| depth-wise conv reduction instead of average pooling in EMSA (ablation) | ResT-Lite · ImageNet-1k | top-1 accuracy: 72.64 -> 72.88 | Helped |
| depth-wise conv reduction instead of max pooling (ablation) | ResT-Lite · ImageNet-1k | top-1 accuracy: 72.20 -> 72.88 | Helped |
| add instance normalization in EMSA (ablation) | ResT-Lite · ImageNet-1k | top-1 accuracy: 71.98 -> 72.88 | Helped |
| add 1x1 head-interaction conv and IN (ablation) | ResT-Lite · ImageNet-1k | top-1 accuracy: 71.72 -> 72.88 | Helped |
| add pixel-attention positional encoding (ablation) | ResT-Lite without PE · ImageNet-1k | top-1 accuracy: 71.54 -> 72.88 | Helped |
| PA positional encoding instead of group linear (ablation) | ResT-Lite · ImageNet-1k | top-1 accuracy: 72.04 -> 72.88 | Helped |
| group linear PE instead of learnable fixed-length PE (ablation) | ResT-Lite · ImageNet-1k | top-1 accuracy: 71.98 -> 72.04 | No clear effect |
Closest papers by meaning
- ResT V2: Simpler, Faster and Stronger (NeurIPS 2022 · similarity 0.92)
- SparseFormer: Sparse Visual Recognition via Limited Latent Tokens (ICLR 2024 · similarity 0.88)
- LookWhere? Efficient Visual Recognition by Learning Where to Look and What to See from Self-Supervision (NeurIPS 2025 · similarity 0.87)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.87)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.86)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.86)
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.