Research map / 3D vision / Human pose estimation
A Dual-Augmentor Framework for Domain Generalization in 3D Human Pose Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a dual-augmentor GAN-style framework (weak and strong 3D pose augmentors with differential generation/discrimination) plus meta-optimization to improve domain generalization for 3D human pose estimation from 2D-to-3D lifting.
Problem
Domain generalization for 3D human pose estimation: pose estimators trained on lab-collected source data (e.g., Human3.6M) generalize poorly to in-the-wild target domains, and existing single-augmentor adversarial methods struggle with unknown target shifts and limited out-of-source exploration.
Approach
Two pose augmentors (weak, which stays close to source; strong, which explores far out-of-source poses) with differential similarity/dissimilarity objectives in generation and a chained WGAN-GP discrimination where weak-augmented poses bridge source and strong poses, plus a MAML-style meta-optimization that simulates domain shift among source, weak, and strong poses during estimator training.
Key results
- On cross-dataset MPI-INF-3DHP, achieves 92.9 PCK, 60.7 AUC, 63.1 MPJPE vs. CEE-Net's 89.9/58.2/69.7.
- On 3DPW cross-dataset, reaches 73.2 PA-MPJPE and 106.6 MPJPE, improving over CEE-Net's 76.8 PA-MPJPE.
- Cross-scenario Human3.6M (S1,S5,S6,S7,S8 to S9,S11) achieves 44.4 MPJPE vs. CEE-Net's 47.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the dual-augmentor framework instead of CEE-Net | VPose (1-frame) · 3DHP test | MPJPE: reduces MPJPE by 6.6mm | Helped |
| use the dual-augmentor framework instead of CEE-Net | VPose (1-frame) · 3DPW | PA-MPJPE: improvement of 3.6mm | Helped |
| add differential generation (ablation) | Dual-augmentor framework · 3DHP test | MPJPE: 5.9mm increase in MPJPE | Helped |
| add differential generation (ablation) | Dual-augmentor framework · 3DPW | MPJPE: 5.2mm | Helped |
| add differential discrimination (ablation) | Dual-augmentor framework · 3DHP test | MPJPE: 4.1mm increase in MPJPE | Helped |
| add differential discrimination (ablation) | Dual-augmentor framework · 3DPW | MPJPE: 3.0mm | Helped |
| add meta-optimization (ablation) | Dual-augmentor framework · 3DHP test | MPJPE: 2.4mm increase in MPJPE | Helped |
| add meta-optimization (ablation) | Dual-augmentor framework · 3DPW | MPJPE: 1.8mm | Helped |
| maintain proximate pairs in the weak augmentor (ablation) | Dual-augmentor framework · 3DHP test | MPJPE: 9.5mm | Helped |
| maintain proximate pairs in the strong augmentor (ablation) | Dual-augmentor framework · 3DHP test | MPJPE: 8.2mm | Helped |
| maintain one-state-gap pairs in the weak augmentor (ablation) | Dual-augmentor framework · 3DHP test | MPJPE: 1.7mm increase in the MPJPE | Helped |
| maintain one-state-gap pairs in the strong augmentor (ablation) | Dual-augmentor framework · 3DHP test | MPJPE: 3.4mm increase in MPJPE | Helped |
Surprising findings
- The strong-augmentor-only variant outperforms the weak-augmentor-only variant, indicating exploring out-of-source distributions matters more than retaining source knowledge in cross-dataset tasks.
Limitations
- Some baselines (CEE-Net, PoseGU) lack released code or full metrics, limiting direct comparison.
- Experiments rely on ground-truth 2D keypoints in main tables with single-frame VPose backbone.
Experimental setup
Datasets: Human3.6M
Benchmarks: MPI-INF-3DHP 3DPW Human3.6M
Metrics: MPJPE PA-MPJPE PCK AUC
Baselines: VPose EvoSkeleton RepNet PoseAug DH-AUG PoseGU CEE-Net VIBE SemGCN PoseFormer MixSTE
Builds on: PoseAug DH-AUG VPose MAML WGAN-GP
Closest papers by meaning
- Lifelong Domain Adaptive 3D Human Pose Estimation (AAAI 2026 · similarity 0.90)
- PoseAug: A Differentiable Pose Augmentation Framework for 3D Human Pose Estimation (CVPR 2021 · similarity 0.90)
- Source-Free and Image-Only Unsupervised Domain Adaptation for Category Level Object Pose Estimation (ICLR 2024 · similarity 0.87)
- ManiPose: Manifold-Constrained Multi-Hypothesis 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.87)
- PoseGen: Learning to Generate 3D Human Pose Dataset with NeRF (AAAI 2024 · similarity 0.86)
- PoseIRM: Enhance 3D Human Pose Estimation on Unseen Camera Settings via Invariant Risk Minimization (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.