Research map / 3D vision / Human pose estimation
PandaPose: 3D Human Pose Lifting from a Single Image via Propagating 2D Pose Prior to 3D Anchor Space
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 PandaPose, a single-image 3D human pose lifting method that propagates 2D pose priors into a 3D anchor space (joint-wise adaptive anchors, joint-wise depth distributions, and an anchor-feature interaction decoder) and predicts joints via anchor-to-joint ensemble prediction.
Problem
Single-image 3D human pose lifting from predicted 2D poses suffers from error propagation from noisy 2D inputs and difficulty resolving self-occlusion and depth ambiguity.
Approach
Instead of direct 2D-to-3D joint regression, build joint-wise adaptive 3D anchors seeded from the 2D pose, lift image features into depth-aware 3D space using joint-wise depth distributions supervised sparsely by GT joint depths, and predict each joint as a weighted sum of anchor-plus-offset positions.
Key results
- On Human3.6M full test set, PandaPose reaches 39.8 mm MPJPE and 32.7 mm PA-MPJPE, versus 41.4 mm and 33.5 mm for CA-PF.
- On the Human3.6M challenging subset, MPJPE drops from 82.4 mm (CA-PF) to 73.1 mm, a 9.3 mm improvement.
- Cross-dataset evaluation on 3DPW (trained on Human3.6M) gives 74.9 mm MPJPE and 46.9 mm PA-MPJPE, beating HiPART by 2.3 and 1.9 mm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PandaPose instead of CA-PF | Image-based 3D pose lifting · Human3.6M full test set | MPJPE (mm): 41.4 -> 39.8 | Helped |
| Use PandaPose instead of CA-PF | Image-based 3D pose lifting · Human3.6M challenging subset | MPJPE (mm): 82.4 -> 73.1 | Helped |
| Use PandaPose instead of CA-PF | Image-based 3D pose lifting · MPI-INF-3DHP full test set | MPJPE (mm): 32.7 -> 31.8 | Helped |
| Use PandaPose instead of HiPART | Cross-dataset 3D pose lifting · 3DPW | MPJPE (mm): 77.2 -> 74.9 | Helped |
| Add global fixed anchors to a model without anchors (ablation) | PandaPose · Human3.6M full test set | MPJPE (mm): 42.1 -> 40.8 | Helped |
| Add adaptive local anchors to a model without anchors (ablation) | PandaPose · Human3.6M full test set | MPJPE (mm): 42.1 -> 40.1 | Helped |
| Combine global fixed and adaptive local anchors (ablation) | PandaPose · Human3.6M full test set | MPJPE (mm): 40.1 -> 39.8 | Helped |
| Replace direct MLP pose regression with anchor-to-joint regression (ablation) | PandaPose · Human3.6M full test set | MPJPE (mm): 42.1 -> 39.8 | Helped |
| Add 3D depth-aware feature lifting and a single depth map (ablation) | PandaPose · Human3.6M full test set | MPJPE (mm): 40.9 -> 40.3 | Helped |
| Replace a single depth map with joint-wise depth distributions (ablation) | PandaPose · Human3.6M challenging subset | MPJPE (mm): 75.9 -> 73.1 | Helped |
| Replace depth regression with classification using 64 bins (ablation) | PandaPose · Human3.6M challenging subset | MPJPE (mm): 77.2 -> 73.1 | Helped |
| Use 128 rather than 64 depth bins (ablation) | PandaPose · Human3.6M full test set | MPJPE (mm): 39.8 -> 40.2 | Hurt |
Surprising findings
- The single-frame PandaPose matches sequence-based SOTA (KTPFormer 40.1 mm) at 39.8 mm MPJPE on Human3.6M without temporal context.
Limitations
- Single-frame design lacks temporal smoothness, causing jitter in continuous motion.
- Added feature lifting and 3D anchor operations increase computational cost.
- Performance depends on the quality of the input 2D pose; failures occur under motion blur or self-occlusion where 2D estimation fails.
Experimental setup
Datasets: Human3.6M MPI-INF-3DHP 3DPW
Benchmarks: Human3.6M MPI-INF-3DHP 3DPW
Metrics: MPJPE PA-MPJPE PCK AUC
Baselines: CA-PF HiPART Di^2Pose DiffPose HCSF GraFormer KTPFormer MixSTE MHFormer PoseFormer
Builds on: A2J-Transformer HRNet Deformable attention
Closest papers by meaning
- Lifting by Image – Leveraging Image Cues for Accurate 3D Human Pose Estimation (AAAI 2024 · similarity 0.92)
- ManiPose: Manifold-Constrained Multi-Hypothesis 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.90)
- Lifting Motion to the 3D World via 2D Diffusion (CVPR 2025 · similarity 0.88)
- COMPOSE: Hypergraph Cover Optimization for Multi-view 3D Human Pose Estimation (NeurIPS 2026 · similarity 0.87)
- Self-Supervised 3D Human Mesh Recovery from a Single Image with Uncertainty-Aware Learning (AAAI 2024 · similarity 0.87)
- VideoMDM: Towards 3D Human Motion Generation From 2D Supervision (NeurIPS 2026 · similarity 0.87)
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.