Research map / Detection and segmentation / CNN features
ResFormer: Scaling ViTs With Multi-Resolution Training
Rui Tian, Zuxuan Wu, Qi Dai, Han Hu, Yu Qiao, Yu-Gang Jiang
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 |
|---|---|---|---|
| Train on multiple resolutions with global-local positional embeddings | ResFormer-T · ImageNet-1K, test 96 | Top-1 accuracy: 8.06 -> 61.40 | Improved |
| Train on multiple resolutions with global-local positional embeddings | ResFormer-S · ImageNet-1K, test 224 | Top-1 accuracy: 79.83 -> 82.16 | Improved |
| Train on multiple resolutions with global-local positional embeddings | ResFormer-B · ImageNet-1K, test 640 | Top-1 accuracy: 74.23 -> 81.72 | Improved |
| Use multi-resolution pretraining | ResFormer-S · ADE20K | mIoU: 45.47 -> 46.31 | Improved |
| Use multi-resolution pretraining | ResFormer-B · COCO2017 mini-val | box AP: +1.3 box AP | Improved |
| Add cross-scale distillation (ablation) | ResFormer-S-MR · ImageNet-1K, test 224 | Top-1 accuracy: 81.77 -> 82.16 | Improved |
| Use one randomly selected resolution per iteration instead of joint multi-resolution batches (ablation) | ResFormer-S-MR · ImageNet-1K, test 224 | Top-1 accuracy: 80.32 -> 82.16 | Hurt |
| Use one resolution per epoch instead of joint multi-resolution batches (ablation) | ResFormer-S-MR · ImageNet-1K, test 128 | Top-1 accuracy: 75.18 -> 78.24 | Hurt |
| Combine global and local positional embeddings (ablation) | ResFormer-S-MR · ImageNet-1K, test 224 | Top-1 accuracy: 81.42 -> 82.16 | Improved |
| Use class-token distillation with smooth L1 rather than logit distillation with KL (ablation) | ResFormer-S-MR · ImageNet-1K, test 224 | Top-1 accuracy: 78.82 -> 79.68 | Improved |
| Use class-token distillation with smooth L1 rather than class-token L2 (ablation) | ResFormer-S-MR · ImageNet-1K, test 224 | Top-1 accuracy: 79.33 -> 79.68 | Improved |
| Apply separate random augmentation at each resolution (ablation) | ResFormer-S-MR · ImageNet-1K, test 128 | Top-1 accuracy: 77.50 -> 78.24 | Improved |
Closest papers by meaning
- Win-Win: Training High-Resolution Vision Transformers from Two Windows (ICLR 2024 · similarity 0.87)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.87)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.87)
- VITRIX-UniViTAR: Unified Vision Transformer with Native Resolution (NeurIPS 2025 · similarity 0.87)
- LF-ViT: Reducing Spatial Redundancy in Vision Transformer for Efficient Image Recognition (AAAI 2024 · similarity 0.87)
- Low-Rank Rescaled Vision Transformer Fine-Tuning: A Residual Design Approach (CVPR 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.
