Research map / 3D vision / Human pose estimation
LiveHPS: LiDAR-based Scene-level Human Pose and Shape Estimation in Free Environment
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 LiveHPS, a single-LiDAR method that jointly estimates SMPL pose, shape and global translation from consecutive point clouds via vertex-guided distillation, attention-based refinement and a skeleton-aware translation solver, and releases the FreeMotion LiDAR motion dataset.
Problem
Scene-level 3D human pose, shape and global translation estimation from LiDAR point clouds in uncontrolled environments, where point distributions vary and occlusion and noise are common.
Approach
Distill pose priors from dense SMPL-vertex point clouds into a LiDAR feature extractor to align varying point distributions, refine joints with temporal and spatial attention, and predict global translation as a bias from the average point location, supervised by a scene-level unidirectional Chamfer distance loss.
Key results
- On FreeMotion, LiveHPS reaches J/V Err(PST) of 119.27/128.61 mm and SUCD of 2.85 mm, versus 129.88/141.77 mm and 4.24 mm for LIP.
- On Sloper4D, LiveHPS achieves J/V Err(PST) of 77.73/85.83 mm, outperforming LIP (94.81/106.34 mm) and LiDARCap.
- Runs in real time at up to 45 fps and maintains reasonable accuracy with short sequences and sparse points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LiveHPS instead of LIP | LiveHPS · SURREAL | SUCD: 0.95 -> 0.67 | Helped |
| Use LiveHPS instead of LIP | LiveHPS · Sloper4D | SUCD: 2.27 -> 1.67 | Helped |
| Use LiveHPS instead of LIP | LiveHPS · FreeMotion | SUCD: 4.24 -> 2.85 | Helped |
| Use LiveHPS instead of LIP | LiveHPS · LiDARHuman26M | SUCD: 3.93 -> 2.67 | Helped |
| Use LiveHPS instead of LIP | LiveHPS · LIPD | SUCD: 2.81 -> 1.98 | Helped |
| Use LiveHPS instead of LIP | LiveHPS · CIMI4D | SUCD: 8.47 -> 4.26 | Helped |
| Use LiveHPS instead of LIP | LiveHPS · SemanticKITTI | SUCD: 9.93 -> 7.28 | Helped |
| Use LiveHPS instead of LIP | LiveHPS · HuCenLife | SUCD: 5.06 -> 3.14 | Helped |
| Add Vertex-guided Adaptive Distillation (ablation) | LiveHPS · FreeMotion | SUCD: 3.17 -> 2.85 | Helped |
| Add Consecutive Pose Optimizer (ablation) | LiveHPS · FreeMotion | SUCD: 4.08 -> 2.85 | Helped |
| Add temporal feature interaction (ablation) | LiveHPS CPO · FreeMotion | SUCD: 3.51 -> 2.85 | Helped |
| Add spatial feature interaction (ablation) | LiveHPS CPO · FreeMotion | SUCD: 2.95 -> 2.85 | Helped |
Surprising findings
- MOVIN's translation estimation diverges badly on FreeMotion and cross-dataset tests, with SUCD in the millions, indicating velocity-based translation accumulates error.
Limitations
- Misjudges global orientation when a human stays static in the scene for a long time, since dynamic cues are lost.
- Relies on a pre-trained segmentation model to isolate human instances at test time.
- Pose and shape errors remain large in far-range or heavily occluded cases (e.g. 100-200 mm-scale joint errors with few points).
Experimental setup
Datasets: FreeMotion Sloper4D SURREAL AIST++ AMASS LiDARHuman26M LIPD CIMI4D SemanticKITTI HuCenLife
Benchmarks: FreeMotion Sloper4D SURREAL LiDARHuman26M LIPD CIMI4D SemanticKITTI HuCenLife
Metrics: J/V Err (mm) Ang Err (deg) SUCD (mm)
Baselines: LiDARCap LIP MOVIN
Builds on: LIP LiDARCap MOVIN PointNet SMPL
Closest papers by meaning
- Pre-training a Density-Aware Pose Transformer for Robust LiDAR-based 3D Human Pose Estimation (AAAI 2025 · similarity 0.90)
- Neighborhood-Enhanced 3D Human Pose Estimation with Monocular LiDAR in Long-Range Outdoor Scenes (AAAI 2024 · similarity 0.88)
- FreeCap: Hybrid Calibration-Free Motion Capture in Open Environments (AAAI 2025 · similarity 0.88)
- RELI11D: A Comprehensive Multimodal Human Motion Dataset and Method (CVPR 2024 · similarity 0.87)
- FreeMan: Towards Benchmarking 3D Human Pose Estimation under Real-World Conditions (CVPR 2024 · similarity 0.85)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 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.