Research map / 3D vision / Human pose estimation
Transformers With Joint Tokens and Local-Global Attention for Efficient Human Pose Estimation
Kaleab A. Kinfu, René Vidal
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 unified skeletal representation and multi-dataset training (ablation) | UniTransPose-B/PS · OCHuman test | mAP: 60.7 -> 93.5 | Improved |
| add unified skeletal representation and multi-dataset training (ablation) | UniTransPose-B/PS · CrowdPose test | mAP: 69.9 -> 78.2 | Improved |
| add unified skeletal representation and multi-dataset training (ablation) | UniTransPose-B/PS · MS-COCO val | mAP: 76.1 -> 78.0 | Improved |
| add JAGL global attention over local cross-window attention (ablation) | UniTransPose-B/PS · MS-COCO val | mAP: 77.1 -> 78.0 | Improved |
| add JAGL global attention over local shifted-window attention (ablation) | UniTransPose-B/PS · OCHuman test | mAP: 89.6 -> 93.5 | Improved |
| use cross-window instead of shifted-window local attention (ablation) | UniTransPose-B/PS · OCHuman test | mAP: 89.6 -> 88.3 | Hurt |
| use pixel-shuffle decoder instead of classical CNN heat-map decoder | UniTransPose-S · MS-COCO val | mAP / FLOPs: 75.9 -> 76.2 mAP, 23.5G -> 10.0G FLOPs | Improved |
| use only last feature map instead of hierarchical maps | UniTransPose-S/PS · MS-COCO val | mAP: 76.2 -> 73.4 | Hurt |
| use key-point regression decoder instead of heat-map decoder | UniTransPose/PS · MS-COCO val | AP: 78.0 -> 68.0 | Hurt |
| add joint-token-based patch selection | EViTPose-B · MS-COCO val | mAP / FLOPs: 77.6 -> 76.5 mAP, 19.8G -> 13.7G FLOPs | Mixed |
| add joint-token patch selection in place of neighboring selection | EViTPose-B · MS-COCO val | mAP: 74.1 -> 76.5 | Improved |
| use hierarchical UniTransPose over ViTPose-H | UniTransPose-B/PS (384x288) · MS-COCO val | mAP: 79.5 -> 79.3 | Mixed |
Closest papers by meaning
- PoseFormerV2: Exploring Frequency Domain for Efficient and Robust 3D Human Pose Estimation (CVPR 2023 · similarity 0.91)
- ViTPose++: Vision Transformer for Generic Body Pose Estimation (TPAMI 2023 · similarity 0.91)
- ExtPose: Robust and Coherent Pose Estimation by Extending ViTs (ICML 2025 · similarity 0.90)
- SDPose: Tokenized Pose Estimation via Circulation-Guide Self-Distillation (CVPR 2024 · similarity 0.88)
- Stable-Pose: Leveraging Transformers for Pose-Guided Text-to-Image Generation (NeurIPS 2024 · similarity 0.88)
- Optimizing Human Pose Estimation Through Focused Human and Joint Regions (AAAI 2025 · similarity 0.88)
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.
