Research map / Detection and segmentation / CNN features
Scale-Aware Modulation Meet Transformer
Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, Lianwen Jin
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 Multi-Head Mixed Convolution in place of single-head convolution (ablation) | attention-free SMT-T · ImageNet-1K val | top-1 accuracy: 80.0 -> 80.8 | Improved |
| add Scale-Aware Aggregation in place of single linear layer (ablation) | SMT-T with MHMC · ImageNet-1K val | top-1 accuracy: 80.8 -> 81.6 | Improved |
| add Evolutionary Hybrid Network (attention in last two stages, hybrid stacking) (ablation) | SMT-T with MHMC and SAA · ImageNet-1K val | top-1 accuracy: 81.6 -> 82.2 | Improved |
| use 4 MHMC heads instead of 1 (ablation) | SMT-T · ImageNet-1K val | top-1 accuracy: 81.8 -> 82.2 | Improved |
| increase MHMC heads beyond 4 (6 or 8) (ablation) | SMT-T · ImageNet-1K val | top-1 accuracy: 82.2 -> 81.9 (6 heads), 82.0 (8 heads) | No clear effect |
| more MHMC heads lowers throughput (ablation) | SMT-T · ImageNet-1K val (V100) | throughput (images/s): 983 -> 833 (1 to 4 heads) | Hurt |
| use SAA instead of no aggregation (ablation) | SMT-T · ImageNet-1K val | top-1 accuracy: 81.5 -> 82.2 | Improved |
| use SAA instead of IBN aggregation (ablation) | SMT-T · ImageNet-1K val | top-1 accuracy: 82.1 -> 82.2 with 12.1M -> 11.5M params | Improved |
| alternate SAM and MSA blocks in penultimate stage rather than SAM-first-half then MSA-second-half (ablation) | SMT-T · ImageNet-1K val | top-1 accuracy: 81.9 -> 82.2 | Improved |
| use hybrid SAM+MSA stacking instead of all-SAM (ablation) | SMT-T · ImageNet-1K val | top-1 accuracy: 81.4 -> 82.2 | Improved |
| use hybrid stacking instead of all-MSA stacking (ablation) | SMT-T · ImageNet-1K val | top-1 accuracy: 81.8 -> 82.2 | Improved |
| SMT-B backbone vs Swin-B in Mask R-CNN 1x | Mask R-CNN · COCO val | box AP: 46.9 -> 49.0 | Improved |
Closest papers by meaning
- MaxViT: Multi-axis Vision Transformer (ECCV 2022 · similarity 0.88)
- A Simple Single-Scale Vision Transformer for Object Detection and Instance Segmentation (ECCV 2022 · similarity 0.88)
- Bridging the Gap Between Vision Transformers and Convolutional Neural Networks on Small Datasets (NeurIPS 2022 · similarity 0.87)
- Frequency-Dynamic Attention Modulation For Dense Prediction (ICCV 2025 · similarity 0.87)
- Scaling Vision Transformers (CVPR 2022 · similarity 0.87)
- Multiscale Vision Transformers (ICCV 2021 · 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.
