Research map / Detection and segmentation / CNN features
ResT V2: Simpler, Faster and Stronger
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 |
|---|---|---|---|
| add pixel-shuffle upsample branch on V to EMSA (ablation) | ResTv2-T · ImageNet-1k 100ep | Top-1 (%): 79.04 -> 80.33 | Helped |
| upsample V instead of x' or K (ablation) | ResTv2-T · ImageNet-1k 100ep | Top-1 (%): 79.64 (x'), 80.03 (K) -> 80.33 (V) | Helped |
| pixel-shuffle instead of nearest/bilinear upsampling (ablation) | ResTv2-T · ImageNet-1k 100ep | Top-1 (%): 79.16 / 79.28 -> 80.33 | Helped |
| nearest or bilinear upsampling vs none (ablation) | ResTv2-T · ImageNet-1k 100ep | Top-1 (%): 79.04 -> 79.16 / 79.28 | Helped |
| replace EMSA with downsample-upsample ConvNet branch only (ablation) | ResTv2-T · ImageNet-1k 100ep | Top-1 (%): 79.04 -> 77.18 | Hurt |
| add positional embedding (PA vs none) (ablation) | ResTv2-T · ImageNet-1k 100ep | Top-1 (%): 79.94 -> 80.33 | Helped |
| use window attention in detection fine-tuning vs global (ablation) | ResTv2-T · COCO Mask R-CNN | AP box: 46.13 -> 43.95 | Hurt |
| use window attention in detection fine-tuning vs global (ablation) | ResTv2-T · COCO Mask R-CNN | ms/iter: 79.9 -> 149.6 | Hurt |
| add multi-head interaction module (head_dim 64) (ablation) | ResTv2-Lite · ImageNet-1k 100ep | Top-1 (%): 74.54 -> 75.04 | Helped |
| add multi-head interaction module (head_dim 96) (ablation) | ResTv2-Tiny · ImageNet-1k 100ep | Top-1 (%): 80.33 -> 80.47 | No clear effect |
| ResTv2-T vs ResTv1-B | ResTv2 · ImageNet-1k | Top-1 (%): 81.6 -> 82.3 | Helped |
| raise resolution from 224 to 384 | ResTv2 · ImageNet-1k | Top-1 (%): 82.3 -> 83.7 (T) | Helped |
Closest papers by meaning
- ResT: An Efficient Transformer for Visual Recognition (NeurIPS 2021 · similarity 0.92)
- MViTv2: Improved Multiscale Vision Transformers for Classification and Detection (CVPR 2022 · similarity 0.87)
- Hiera: A Hierarchical Vision Transformer without the Bells-and-Whistles (ICML 2023 · similarity 0.85)
- LookWhere? Efficient Visual Recognition by Learning Where to Look and What to See from Self-Supervision (NeurIPS 2025 · similarity 0.84)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.84)
- ImagePiece: Content-aware Re-tokenization for Efficient Image Recognition (AAAI 2025 · similarity 0.84)
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.