Research map / Label-efficient and robust learning / Domain adaptation and generalization
HamiPose: Hamiltonian Optimization for Unsupervised Domain Adaptive Pose Estimation
Jiawen Li, Fei Jiang, Dandan Zhu, Aimin Zhou
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 metric (model-wide mass with head-level normalizer) (ablation) | Mean Teacher baseline · SURREAL->Human3.6M | PCK@0.05 All: 75.3 -> 77.2 | Improved |
| add unified metric (ablation) | Mean Teacher baseline · RHD->H3D | PCK@0.05 All: 79.3 -> 80.9 | Improved |
| add channelwise gated filtering on top of unified metric (ablation) | Mean Teacher + unified metric · SURREAL->Human3.6M | PCK@0.05 All: 77.2 -> 78.5 | Improved |
| add channelwise gated filtering on top of unified metric (ablation) | Mean Teacher + unified metric · RHD->H3D | PCK@0.05 All: 80.9 -> 82.3 | Improved |
| add Hamiltonian symplectic transport (ablation) | Mean Teacher + metric + filtering · SURREAL->Human3.6M | PCK@0.05 All: 78.5 -> 79.8 | Improved |
| add Hamiltonian symplectic transport (ablation) | Mean Teacher + metric + filtering · RHD->H3D | PCK@0.05 All: 82.3 -> 83.1 | Improved |
| use keypointwise gating instead of unified gating (ablation) | HamiPose gradient filter · UDA training | source-target gradient conflict ratio at epoch 30: 8 to 9% -> 6% | Improved |
| use Hamiltonian optimizer instead of SGD or Adam (ablation) | HamiPose training · UDA training loss | training loss stability | Mixed |
| use moderate gamma_max (2.0) with 10% warmup ratio, gamma_min 0 (ablation) | channelwise gate · RHD->H3D | PCK@0.05: 68.0 -> 70.3 (gamma_max 1.0 5% warmup vs 2.0 10% warmup) | Improved |
| start gate sharpness gamma_min at 0 instead of 0.5 (ablation) | channelwise gate · RHD->H3D | PCK@0.05: 69.7 -> 70.3 | Improved |
| use overly long warmup ratio (20%) instead of 10% (ablation) | channelwise gate · RHD->H3D | PCK@0.05: 70.3 -> 69.5 | Hurt |
| apply HamiPose to domain generalization (train SURREAL->LSP, test Human3.6M) | HamiPose · Human3.6M (unseen) | PCK@0.05 All: 76.8 -> 77.5 vs GGA | Improved |
Closest papers by meaning
- A Unified Framework for Domain Adaptive Pose Estimation (ECCV 2022 · similarity 0.87)
- A Dual-Augmentor Framework for Domain Generalization in 3D Human Pose Estimation (CVPR 2024 · similarity 0.86)
- Source-Free and Image-Only Unsupervised Domain Adaptation for Category Level Object Pose Estimation (ICLR 2024 · similarity 0.85)
- ConDo: Continual Domain Expansion for Absolute Pose Regression (AAAI 2025 · similarity 0.85)
- Fast Adaptation for Human Pose Estimation via Meta-Optimization (CVPR 2024 · similarity 0.84)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.84)
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.
