Research map / 3D vision / Human pose estimation
Toward Approaches to Scalability 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 Biomechanical Pose Generator that synthesizes plausible 3D poses from normal range of motion constraints and a Binary Depth Coordinates formulation that reduces depth lifting to a front/back binary choice, improving 3D HPE generalization on H36M, 3DHP and 3DPW.
Problem
Scalability of 3D human pose estimation is limited by popularity bias in existing datasets and by one-to-many depth ambiguity that grows as pose diversity increases.
Approach
Generate out-of-source 3D poses from an osteo-kinematic model constrained by normal range of motion, bone length ratios and pose confidence, and decompose each 3D pose into 2D pose, bone length and a binary depth decision (solved via a quadratic, with Gumbel-Softmax for grouped segments).
Key results
- BPG + BDC reaches 44.0 MPJPE and 34.3 P-MPJPE on H36M (S9, S11 targets) versus 52.7/40.6 for VPose and 47.3/36.8 for CEE-Net.
- On 3DPW cross-dataset evaluation, BPG + BDC reduces P-MPJPE to 61.1 from 94.6 for VPose and 76.8 for CEE-Net.
- On MPI-INF-3DHP, BPG + BDC achieves 91.8 PCK and 61.8 MPJPE, versus 89.9 PCK and 69.7 MPJPE for CEE-Net.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BPG pose augmentation to VPose | VPose · H36M full training | MPJPE: 52.7 -> 46.9 | Helped |
| Add BDC with BPG pose augmentation | VPose · H36M full training | MPJPE: 46.9 -> 44.0 | Helped |
| Add BPG pose augmentation and BDC | VPose · 3DHP | MPJPE: 69.7 -> 61.8 | Helped |
| Add BPG pose augmentation and BDC | VPose · 3DPW | P-MPJPE: 76.8 -> 61.1 | Helped |
| Add NROM (ablation) | BPG kinematic model · H36M BPG-only training | MPJPE: 122.3 -> 65.7 | Helped |
| Add rotational restriction to Variant A (ablation) | BPG · H36M BPG-only training | MPJPE: 65.7 -> 64.2 | Helped |
| Add NROM and body ratio variation without rotational restriction (ablation) | BPG · H36M BPG-only training | MPJPE: 122.3 -> 62.6 | Helped |
| Add rotational restriction to Variant C (ablation) | BPG · H36M BPG-only training | MPJPE: 62.6 -> 61.9 | Helped |
| Add pose confidence to Variant D (ablation) | BPG · H36M BPG-only training | MPJPE: 61.9 -> 59.8 | Helped |
| Replace CSC with NDC (ablation) | GraphMDN · H36M (S9, S11 target) | MPJPE: 46.2 -> 44.3 | Helped |
| Replace CSC with BDC without grouped segments (ablation) | GraphMDN · H36M (S9, S11 target) | MPJPE: 46.2 -> 42.6 | Helped |
| Add grouped segments to BDC (ablation) | GraphMDN · H36M (S9, S11 target) | MPJPE: 42.6 -> 41.7 | Helped |
Surprising findings
- Adding BPG-generated poses without the BDC degrades source-dataset performance as data volume grows, while the BDC reverses this trend.
- Introducing rotational restriction in the BPG reduced performance in ablation (Variants A and C).
Limitations
- BPG generates single-frame poses only, missing temporal dynamics.
- Generating additional data beyond a certain point yields no further improvement (saturation).
- Performance of BPG-only augmentation on source data degrades in some scenarios per the appendix analysis.
Experimental setup
Datasets: Human3.6M MPI-INF-3DHP 3DPW
Benchmarks: Human3.6M MPI-INF-3DHP 3DPW
Metrics: MPJPE P-MPJPE PCK AUC
Baselines: VPose EvoSkeleton PoseAug DH-AUG CEE-Net PoseGU VIBE RepNet GraphMDN GFpose Conpose MixSTE
Builds on: VPose MixSTE GraphMDN PoseAug Drover et al. projection
Closest papers by meaning
- ProPose: Probabilistic 3D Human Pose Estimation with Instance-Level Distribution and Normalizing Flow (AAAI 2025 · similarity 0.87)
- Reconstructing Humans with a Biomechanically Accurate Skeleton (CVPR 2025 · similarity 0.87)
- Lifting by Image – Leveraging Image Cues for Accurate 3D Human Pose Estimation (AAAI 2024 · similarity 0.86)
- PandaPose: 3D Human Pose Lifting from a Single Image via Propagating 2D Pose Prior to 3D Anchor Space (NeurIPS 2025 · similarity 0.86)
- ManiPose: Manifold-Constrained Multi-Hypothesis 3D Human Pose Estimation (NeurIPS 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.