Research map / 3D vision / Human pose estimation
Full-Body Motion Reconstruction with Sparse Sensing from Graph Perspective
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 Body Pose Graph (BPG) framework that treats full-body motion reconstruction from three VR-tracked joints as missing-node prediction, using temporal pyramid features and a GCN with learned expressive edges.
Problem
Reconstructing full-body 3D pose and motion from sparse VR sensor signals (one headset and two hand controllers), where 19 of 22 joints are unobserved and lower-body estimation is especially poor.
Approach
Model the body as a 22-node graph where only three head/hand nodes are observed, initialize node features with temporal pyramid and trunk/limb spatial features, and update them with a GCN whose adjacency combines static skeleton, dynamic skeleton, and latent learned edges.
Key results
- On AMASS, the method reaches MPJRE 2.49, MPJPE 3.34 cm and MPJVE 22.84 cm/s, beating AvatarPoser retrained locally (3.01 / 4.11 / 27.79).
- Lower-body joint position errors are 18-22% lower than AvatarPoser across the listed lower-body joints.
- Against AGRoL with 41-frame input, the method is better on MPJRE and MPJPE (2.49 vs 2.59, 3.34 vs 3.64) but not on MPJVE (22.84 vs 23.24 for AGRoL41 is close; AGRoL196 reaches 18.59).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed method instead of AvatarPoser* | Full-body motion reconstruction · AMASS | MPJPE: 4.11 -> 3 . 3 4 | Helped |
| use the proposed method instead of AGRoL 41 | Full-body motion reconstruction · AMASS, offline comparison | MPJPE: 3.64 -> 3.34 | Helped |
| extend AGRoL output sequence length from 41 to 196 | AGRoL · AMASS, offline comparison | MPJVE: 23.24 -> 18.59 | Helped |
| add bone symmetric loss (ablation) | Full-body motion reconstruction · AMASS | MPJPE-lower-body: 6.75 -> 6.29 | Helped |
| add spatial property generation and trunk-limb relations (ablation) | Full-body motion reconstruction · AMASS | MPJPE-lower-body: 6.88 -> 6.29 | Helped |
| use Temporal Pyramid Structure instead of a temporal feature extractor (ablation) | Full-body motion reconstruction · AMASS | MPJPE-lower-body: 7.10 -> 6.29 | Helped |
| use Feature Integration instead of a simple MLP (ablation) | Full-body motion reconstruction · AMASS | MPJPE-lower-body: 6.80 -> 6.29 | Helped |
| use GCN with expressive edges instead of Vanilla GCN (ablation) | Full-body motion reconstruction · AMASS | MPJPE-lower-body: 6.88 -> 6.29 | Helped |
Limitations
- Lags AGRoL with 196-frame output on MPJVE (18.59 vs 22.84), so velocity smoothness is weaker.
- Baseline results for several methods are taken from prior papers rather than re-run, and AvatarPoser was retrained only by the authors.
- Evaluated only on AMASS-derived sets; generalization to diverse real VR recordings is not shown.
Experimental setup
Datasets: CMU BMLrub HDM05 AMASS
Benchmarks: AMASS
Metrics: MPJRE MPJPE MPJVE MPJPE-lower-body
Baselines: Final IK CoolMoves LoBSTr VAE-HMD AvatarPoser AGRoL
Builds on: AvatarPoser SCI-Block graph convolutional network SMPL
Closest papers by meaning
- A Unified Diffusion Framework for Scene-aware Human Motion Estimation from Sparse Signals (CVPR 2024 · similarity 0.85)
- Spatial-Related Sensors Matters: 3D Human Motion Reconstruction Assisted with Textual Semantics (AAAI 2024 · similarity 0.85)
- EnvPoser: Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling (CVPR 2025 · similarity 0.84)
- Stratified Avatar Generation from Sparse Observations (CVPR 2024 · similarity 0.84)
- Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors (CVPR 2024 · similarity 0.83)
- DiffusionPoser: Real-time Human Motion Reconstruction From Arbitrary Sparse Sensors Using Autoregressive Diffusion (CVPR 2024 · similarity 0.83)
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.