Research map / 3D vision / Human pose estimation
Neighborhood-Enhanced 3D Human Pose Estimation with Monocular LiDAR in Long-Range Outdoor Scenes
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 LiDAR-based 3D human pose estimation method that adds background neighborhood points (3BN) and 3D scanning neighbor points (3SN) fused by a transformer CoherenceFuse module, evaluated on LidarHuman26M, CIMI4D, SLOPER4D and Waymo.
Problem
Monocular LiDAR 3D human pose estimation in long-range outdoor scenes suffers from sparse, degraded human point clouds at distance, and existing methods ignore coherence cues in the surrounding neighborhood.
Approach
Exploit environment points near the human (3BN, spatial cue) and the LiDAR scanning neighbor points lying behind the human silhouette (3SN, structural edge cue), fusing them with human points via a transformer module (CoherenceFuse) built on a LidarCap-style kinematics pipeline.
Key results
- On LidarHuman26M, MPJPE drops from 79.31mm (LidarCap) to 72.23mm, a 7.08mm reduction.
- For targets beyond 25m, MPJPE improves by about 16.55mm over LidarCap on LidarHuman26M.
- On CIMI4D MPJPE improves by 8.66mm and on SLOPER4D by 5.9mm relative to LidarCap; on Waymo chamfer distance improves 4.79mm over LidarCap.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add 3BN neighborhood features (ablation) | LidarCap · LidarHuman26M | MPJPE (mm): 79.31 -> 75.30 | Helped |
| replace 3BN with 3SN features (ablation) | LidarCap with 3BN · LidarHuman26M | MPJPE (mm): 75.30 -> 75.99 | Hurt |
| add separate 3BN and 3SN encoders (ablation) | LidarCap · LidarHuman26M | MPJPE (mm): 79.31 -> 74.67 | Helped |
| stack the three input branches and use one PointNet++ encoder (ablation) | LidarCap · LidarHuman26M | MPJPE (mm): 79.31 -> 77.87 | Hurt |
| replace GRU with multi-head self-attention (ablation) | LidarCap with 3BN and 3SN inputs · LidarHuman26M | MPJPE (mm): 79.31 -> 73.68 | Helped |
| replace GRU with CoherenceFuse (ablation) | LidarCap with 3BN and 3SN inputs · LidarHuman26M | MPJPE (mm): 79.31 -> 72.23 | Helped |
| increase 3BN radius from 2m to 3m (ablation) | 3BN model · LidarHuman26M | MPJPE: 74.47 -> 76.31 | Hurt |
| increase 3BN radius from 2m to 5m (ablation) | 3BN model · LidarHuman26M | MPJPE: 74.47 -> 76.03 | Hurt |
| increase 3BN radius from 2m to 10m (ablation) | 3BN model · LidarHuman26M | MPJPE: 74.47 -> 75.24 | Hurt |
| increase 3BN radius from 2m to 20m (ablation) | 3BN model · LidarHuman26M | MPJPE: 74.47 -> 75.71 | Hurt |
| add 3BN and 3SN neighborhood cues | LidarCap · LidarHuman26M | MPJPE (mm): 79.31 -> 72.23 | Helped |
| add 3BN and 3SN neighborhood cues | LidarCap · CIMI4D | MPJPE (mm): 130.43 -> 121.77 | Helped |
Surprising findings
- Shadow points computed from 3BN plane projection are more effective than 3SN at close range, but 3SN outperforms shadows at long distances.
Limitations
- Relies on ideal foreground human point clouds; no segmentation or detection over full frames.
- Inferior to HSC2 on acceleration error since HSC2 uses post-optimization physical constraints.
- Gains from background radius are limited; larger background coverage does not help.
Experimental setup
Datasets: LidarHuman26M CIMI4D SLOPER4D Waymo Open Dataset v2.0
Benchmarks: LidarHuman26M CIMI4D SLOPER4D Waymo Open Dataset v2.0
Metrics: MPJPE PA-MPJPE PCK30 PCK50 acceleration error chamfer distance
Baselines: LidarCap P4T HSC2 CLIFF
Builds on: LidarCap PointNet++ ST-GCN 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)
- LiveHPS: LiDAR-based Scene-level Human Pose and Shape Estimation in Free Environment (CVPR 2024 · similarity 0.88)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.85)
- FreeMan: Towards Benchmarking 3D Human Pose Estimation under Real-World Conditions (CVPR 2024 · similarity 0.85)
- USCILab3D: A Large-scale, Long-term, Semantically Annotated Outdoor Dataset (NeurIPS 2024 · similarity 0.85)
- Mind The Edge: Refining Depth Edges in Sparsely-Supervised Monocular Depth Estimation (CVPR 2024 · similarity 0.85)
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.