Research map / Point clouds / Point cloud classification
LidarGait++: Learning Local Features and Size Awareness from LiDAR Point Clouds for 3D Gait Recognition
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 LidarGait++, a point-based 3D gait recognition network using set abstraction layers, a pyramid point pooling layer, and size-aware encoding/partitioning that injects absolute altitude, reaching 92.7% rank-1 on SUSTech1K and 82.0% on FreeGait.
Problem
Point-based deep networks underperform projection-based methods for 3D gait recognition from LiDAR point clouds, and existing methods normalize away body size cues that help distinguish subjects.
Approach
Align point-based architectures with grid-based gait design (local feature learning via set abstraction and horizontal-pyramid-style pooling), and add size awareness by feeding absolute altitude into the encoder and using a highest-altitude partition that keeps subjects at their true scale.
Key results
- LidarGait++ reaches 92.7% rank-1 on SUSTech1K overall, versus 86.8% for LidarGait and 90.2% for HMRNet, and 77.1% for the PointNet++ baseline.
- On FreeGait self-domain, LidarGait++ achieves 82.0% rank-1 and 87.2% mAP, outperforming HMRNet (80.8%) and LidarGait (74.2%).
- In cross-domain SUSTech1K-to-FreeGait, LidarGait++ gets 15.1% rank-1, slightly below LidarGait's 23.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add size-aware encoding layer (ablation) | PointNet++ · SUSTech1K overall | Rank-1 accuracy: 77.1% to 80.8% | Helped |
| add size-aware encoding layer (ablation) | PointNet++ · SUSTech1K clothes-changing subset | Rank-1 accuracy: 58.7% to 70.7% | Helped |
| add naive P3 with Uniform Density Partition (ablation) | size-aware PointNet++ · SUSTech1K overall | Rank-1 accuracy: 80.8 to 88.6 | Helped |
| use Uniform Altitude Partition instead of Uniform Density Partition in P3 (ablation) | size-aware PointNet++ with P3 · SUSTech1K overall | Rank-1 accuracy: 88.6 to 89.3 | Helped |
| use size-aware Highest Altitude Partition in P3 instead of Uniform Altitude Partition (ablation) | size-aware PointNet++ with P3 · SUSTech1K overall | Rank-1 accuracy: 89.3 to 92.7 | Helped |
| replace P3 with Global Max Pooling (ablation) | PointNet++ · SUSTech1K overall | Rank-1 accuracy: 92.7 to 80.8 | Hurt |
| replace P3 with NetVLAD (ablation) | PointNet++ · SUSTech1K overall | Rank-1 accuracy: 92.7 to 85.8 | Hurt |
| use width as an additional size prior (ablation) | LidarGait++ · SUSTech1K | Rank-1 accuracy: +width (↑1.2%) | Helped |
| use depth as an additional size prior (ablation) | LidarGait++ · SUSTech1K | Rank-1 accuracy: +depth (↑1.6%) | Helped |
| use height as a size prior (ablation) | LidarGait++ · SUSTech1K | Rank-1 accuracy: +height (↑8.1%) | Helped |
| use height, width, and depth as size priors (ablation) | LidarGait++ · SUSTech1K | Rank-1 accuracy: +height+width+depth (↑8.5%) | Helped |
| replace convolutional kernels of size 3×3 with smaller kernels (ablation) | LidarGait · SUSTech1K | Rank-1 accuracy | Hurt |
Surprising findings
- Cross-domain generalization from SUSTech1K to FreeGait is weaker for LidarGait++ than for projection-based LidarGait.
- Height is the strongest single prior: adding height gives +8.1% while width and depth give only +1.2% and +1.6%.
Limitations
- Cross-domain performance from SUSTech1K to FreeGait is weaker than projection-based LidarGait.
- No temporal modeling design in the model.
- Set abstraction encoding is computationally inefficient.
Experimental setup
Datasets: SUSTech1K FreeGait
Benchmarks: SUSTech1K FreeGait
Metrics: rank-1 accuracy rank-5 accuracy mAP
Baselines: GaitSet GaitGL GaitBase PointNet PointTransformer DGCNN PointGait SimpleView LidarGait HMRNet PointNet++ PointMLP
Builds on: PointNet++ LidarGait Horizontal Pyramid Pooling OpenGait
Closest papers by meaning
- LidarGait: Benchmarking 3D Gait Recognition With Point Clouds (CVPR 2023 · similarity 0.94)
- BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models (NeurIPS 2025 · similarity 0.86)
- BigGait: Learning Gait Representation You Want by Large Vision Models (CVPR 2024 · similarity 0.85)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.84)
- iDet3D: Towards Efficient Interactive Object Detection for LiDAR Point Clouds (AAAI 2024 · similarity 0.84)
- HybridGait: A Benchmark for Spatial-Temporal Cloth-Changing Gait Recognition with Hybrid Explorations (AAAI 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.