Research map / Architectures and efficiency / Attention and transformers
TopFormer: Token Pyramid Transformer for Mobile Semantic Segmentation
Wenqiang Zhang, Zilong Huang, Guozhong Luo, Tao Chen, Xinggang Wang, Wenyu Liu, Gang Yu, Chunhua Shen
Check 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 |
|---|---|---|---|
| use multi-scale token pyramid rather than only the last-scale token as Semantics Extractor input (ablation) | TopFormer-B · ADE20K val | mIoU: 36.6 -> 37.8 | Improved |
| inject semantics at scales 1/4, 1/8, 1/16, and 1/32 rather than at 1/8, 1/16, and 1/32 (ablation) | TopFormer-B · ADE20K val | mIoU: 37.8 -> 38.3 | Improved |
| add injection at the 1/8 scale to the 1/16 and 1/32 scales (ablation) | TopFormer-B · ADE20K val | mIoU: 36.4 -> 37.8 | Improved |
| add the Scale-aware Semantics Extractor (ablation) | TopFormer-T · ADE20K val | mIoU: 22.5 -> 32.8 | Improved |
| replace MHSA modules with additional FFNs (ablation) | TopFormer-T · ADE20K val | mIoU: 32.8 -> 30.4 | Hurt |
| use SASE instead of PSP contextual modeling (ablation) | TopFormer-T · ADE20K val | mIoU: 27.4 -> 32.8 | Improved |
| use SASE instead of ASPP contextual modeling (ablation) | TopFormer-T · ADE20K val | mIoU: 28.2 -> 32.8 | Improved |
| combine SigmoidAttn with SemInfo rather than using either alone (ablation) | TopFormer · ADE20K val | mIoU: 37.3 -> 37.8; 37.0 -> 37.8 | Improved |
| use the proposed segmentation head instead of Concat Head (ablation) | TopFormer · ADE20K val | mIoU: 37.1 -> 37.8 | Improved |
| use the proposed segmentation head instead of Sum Head (ablation) | TopFormer · ADE20K val | mIoU: 37.0 -> 37.8 | Improved |
| increase SIM width from 128 to 192 or 256 (ablation) | TopFormer · ADE20K val | mIoU: 37.3 -> 37.4; 37.3 -> 37.8 | Improved |
| increase pooled-token output stride from s128 to s64 (ablation) | TopFormer · ADE20K val | mIoU: 35.7 -> 37.8 | Improved |
Closest papers by meaning
- SeaFormer: Squeeze-enhanced Axial Transformer for Mobile Semantic Segmentation (ICLR 2023 · similarity 0.90)
- FeedFormer: Revisiting Transformer Decoder for Efficient Semantic Segmentation (AAAI 2023 · similarity 0.88)
- SeaFormer++: Squeeze-Enhanced Axial Transformer for Mobile Visual Recognition (IJCV 2025 · similarity 0.87)
- Token Cropr: Faster ViTs for Quite a Few Tasks (CVPR 2025 · similarity 0.86)
- DTMFormer: Dynamic Token Merging for Boosting Transformer-Based Medical Image Segmentation (AAAI 2024 · similarity 0.85)
- ALGM: Adaptive Local-then-Global Token Merging for Efficient Semantic Segmentation with Plain Vision Transformers (CVPR 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.
