Research map / Detection and segmentation / CNN features
TCFormer: Visual Recognition via Token Clustering Transformer
Wang Zeng, Sheng Jin, Lumin Xu, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, Xiaogang Wang
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 global CTM token clustering in place of CNN merge (ablation) | TCFormerV2-Small + Semantic FPN · ADE20K val | mIoU: 45.2 -> 46.1 | Improved |
| add local CTM token clustering in place of CNN merge (ablation) | TCFormerV2-Small + Semantic FPN · ADE20K val | mIoU: 45.2 -> 45.7 | Improved |
| use local instead of global CTM (ablation) | TCFormerV2-Small + Semantic FPN · ADE20K val | GFLOPs of CTM: 42.1 GFLOPs -> 5.6 GFLOPs | Improved |
| replace FPN with SR-MTA aggregation (local CTM) (ablation) | TCFormerV2-Small + Semantic FPN · ADE20K val | mIoU: 45.7 -> 47.1 | Improved |
| use CR-MTA instead of SR-MTA (local CTM) (ablation) | TCFormerV2-Small + Semantic FPN · ADE20K val | mIoU: 47.1 -> 47.8 | Improved |
| use CR-MTA instead of SR-MTA (global CTM) (ablation) | TCFormerV2-Small + Semantic FPN · ADE20K val | mIoU: 46.5 -> 47.0 | Improved |
| swap local DPC-kNN for K-means at test time without fine-tuning (ablation) | TCFormerV2-Small · ImageNet-1K val | top-1 accuracy: 82.40 -> 82.39 | No clear effect |
| swap local DPC-kNN for Bipartite Soft Matching at test time (ablation) | TCFormerV2-Small · ImageNet-1K val | top-1 accuracy: 82.40 -> 82.31 | No clear effect |
| use local instead of global clustering (TCFormerV2 vs V1) | TCFormerV2-Small · ImageNet-1K val | GFLOPs: 5.9 -> 4.5 | Improved |
| use dynamic-token TCFormerV2-Small instead of grid-token HRNet-w32 | COCO-WholeBody · COCO-WholeBody V1.0 val | hand keypoint AP: 0.473 -> 0.535 | Improved |
| use TCFormerV2-Small instead of PVTv2-B2 for small objects | RetinaNet 1x · COCO val2017 | AP_S: 27.4 -> 28.8 | Improved |
| use TCFormerV2-Base instead of Swin-B with Mask2Former | Mask2Former 640x640 · ADE20K val | mIoU: 52.4 -> 53.8 | Improved |
Closest papers by meaning
- Not All Tokens Are Equal: Human-Centric Visual Analysis via Token Clustering Transformer (CVPR 2022 · similarity 0.92)
- DTMFormer: Dynamic Token Merging for Boosting Transformer-Based Medical Image Segmentation (AAAI 2024 · similarity 0.87)
- ImagePiece: Content-aware Re-tokenization for Efficient Image Recognition (AAAI 2025 · similarity 0.87)
- Token Cropr: Faster ViTs for Quite a Few Tasks (CVPR 2025 · similarity 0.85)
- Not All Tokens Matter All The Time: Dynamic Token Aggregation Towards Efficient Detection Transformers (ICML 2025 · similarity 0.85)
- TGBFormer: Transformer-GraphFormer Blender Network for Video Object Detection (AAAI 2025 · 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.
