Research map / Detection and segmentation / CNN features
SeaFormer++: Squeeze-Enhanced Axial Transformer for Mobile Visual Recognition
Qiang Wan, Zilong Huang, Jiachen Lu, Gang Yu, Li Zhang
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 detail enhancement kernel to squeeze axial attention (ablation) | SeaFormer-T · ADE20K val | mIoU: 33.5 -> 35.8 | Improved |
| use multiplication instead of addition for enhancement mode (ablation) | SeaFormer-T · ADE20K val | mIoU: 35.4 -> 35.8 | Improved |
| share convolution weights between detail enhancement and squeeze attention (ablation) | SeaFormer-T · ADE20K val | mIoU: 35.9 -> 35.8 | No clear effect |
| use adaptive squeeze and expand instead of mean pooling and broadcast (ablation) | SeaFormer-T · ImageNet-1K and ADE20K | top-1 / mIoU: 67.9 -> 69.8 top-1; 35.8 -> 36.8 mIoU | Improved |
| use max pooling instead of mean pooling for squeeze (ablation) | SeaFormer-T · ADE20K val | mIoU: 35.8 -> 35.0 | Hurt |
| use SEA attention instead of global attention (ablation) | SeaFormer-T · ImageNet-1K and ADE20K | top-1 / mIoU: +1.2% Top1; +1.6 mIoU | Improved |
| replace window attention in Swin with SEA attention (ablation) | Swin · ADE20K val | mIoU: 45.3 -> 46.5 | Improved |
| use Light head instead of Simple head (ablation) | SeaFormer-B · ADE20K val | mIoU: +0.4 mIoU | Improved |
| use MobileNetV2-based upsample module instead of direct interpolation (ablation) | SeaFormer++ KD student · ADE20K val | mIoU: 33.7 -> 35.5 | Improved |
| use standard conv upsample instead of MobileNetV2 module (ablation) | SeaFormer++ KD student · ADE20K val | mIoU / latency: 35.5 -> 35.7; 22ms -> 30ms | Mixed |
| add output similarity loss to classification loss in KD (ablation) | SeaFormer++ KD · ADE20K val | mIoU: 32.5 -> 33.7 | Improved |
| add feature similarity loss (ablation) | SeaFormer++ KD · ADE20K val | mIoU: 33.7 -> 34.7 | Improved |
Closest papers by meaning
- SeaFormer: Squeeze-enhanced Axial Transformer for Mobile Semantic Segmentation (ICLR 2023 · similarity 0.97)
- TopFormer: Token Pyramid Transformer for Mobile Semantic Segmentation (CVPR 2022 · similarity 0.87)
- DualFormer: Local-Global Stratified Transformer for Efficient Video Recognition (ECCV 2022 · similarity 0.86)
- Mobile-Former: Bridging MobileNet and Transformer (CVPR 2022 · similarity 0.86)
- FeedFormer: Revisiting Transformer Decoder for Efficient Semantic Segmentation (AAAI 2023 · similarity 0.86)
- Mobile Attention: Mobile-Friendly Linear-Attention for Vision Transformers (ICML 2024 · similarity 0.85)
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.
