Research map / Architectures and efficiency / Attention and transformers
ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation
Yufei Xu, Jing Zhang, Qiming Zhang, Dacheng Tao
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 |
|---|---|---|---|
| scale the backbone from ViT-B to ViT-L | ViTPose · MS COCO val | AP: 75.8 AP v.s 78.3 AP | Improved |
| use a plain ViTAE-G backbone in a single model | ViTPose · MS COCO test-dev | AP: 80.9 AP; 80.8 AP | Improved |
| replace the classic decoder with the simple decoder (ablation) | ViTPose-B, ViTPose-L, and ViTPose-H · MS COCO val | AP: 75.8 -> 75.5; 78.3 -> 78.2; 79.1 -> 78.9 | Hurt |
| replace the classic decoder with the simple decoder (ablation) | SimpleBaseline with ResNet-50 and ResNet-152 · MS COCO val | AP: 71.8 -> 53.1; 73.5 -> 55.3 | Hurt |
| pre-train on COCO crops instead of ImageNet-1k (ablation) | ViTPose-B · MS COCO val | AP: 75.8 -> 74.5 | Hurt |
| pre-train on COCO+AI Challenger data instead of ImageNet-1k (ablation) | ViTPose-B · MS COCO val | AP: 75.8 -> 75.8 | No clear effect |
| increase input resolution from 256x192 to 576x432 (ablation) | ViTPose-B · MS COCO val | AP: 75.8 -> 77.8 | Improved |
| use unshifted, unpooled window attention instead of full attention (ablation) | ViTPose-B at 1/8 feature size · MS COCO val | AP: 77.4 -> 66.4 | Hurt |
| add shifted windows to window attention (ablation) | ViTPose-B at 1/8 feature size · MS COCO val | AP: 66.4 -> 76.4 | Improved |
| add pooled window tokens to window attention (ablation) | ViTPose-B at 1/8 feature size · MS COCO val | AP: 66.4 -> 76.4 | Improved |
| combine shifted windows with pooled window tokens (ablation) | ViTPose-B at 1/8 feature size · MS COCO val | AP: 76.4 -> 76.8 | Improved |
| freeze MHSA modules during finetuning (ablation) | ViTPose-B · MS COCO val | AP: 75.8 -> 75.1 | Hurt |
Closest papers by meaning
- ViTPose++: Vision Transformer for Generic Body Pose Estimation (TPAMI 2023 · similarity 0.96)
- Beyond Static Frames: Temporal Aggregate-and-Restore Vision Transformer for Human Pose Estimation (CVPR 2026 · similarity 0.90)
- ExtPose: Robust and Coherent Pose Estimation by Extending ViTs (ICML 2025 · similarity 0.89)
- SHaRPose: Sparse High-Resolution Representation for Human Pose Estimation (AAAI 2024 · similarity 0.88)
- SDPose: Tokenized Pose Estimation via Circulation-Guide Self-Distillation (CVPR 2024 · similarity 0.86)
- Multiple View Geometry Transformers for 3D Human Pose Estimation (CVPR 2024 · similarity 0.86)
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.
