Research map / Point clouds / Shape and surface completion
LiSu: A Dataset and Method for LiDAR Surface Normal Estimation
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.
Introduces LiSu, a 50k-frame synthetic CARLA LiDAR dataset with surface normal labels, and a Point Transformer V3 method with spatial/temporal graph total variation and Eikonal regularizers that enables LiDAR normal estimation and self-training domain adaptation to Waymo.
Problem
Surface normal estimation from LiDAR point clouds in autonomous driving is underexplored, due to the lack of large annotated datasets and methods that handle sparse, noisy LiDAR efficiently.
Approach
Extend CARLA's LiDAR ray tracer to output ground-truth surface normals, and train a single-pass PTv3 normal estimator with spatial GTV, temporal GTV across consecutive frames, and Eikonal regularization as data terms, which also suppress pseudo-label noise in self-training.
Key results
- On the LiSu test split, the proposed method reaches 6.30° mean and 0.43° median angular error with 81.10% accuracy at 5°, versus 6.46° mean and 79.66% for SHS-Net.
- Inference takes about 0.09 s per frame versus over 14 s for PCPNet, GraphFit and SHS-Net.
- In Waymo neural surface reconstruction, the full regularized self-training model lowers average RMSE/CD from 2.01/0.41 (ReSimAD) to 1.97/0.37; LL1-only self-training gives 2.00/0.40.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add spatial GTV regularization (ablation) | PTv3 surface normal estimator · LiSu in-domain (20% test split) | RMSE angular error (°): 19.82 -> 17.78 | Helped |
| Add temporal GTV regularization (ablation) | PTv3 surface normal estimator · LiSu in-domain (20% test split) | RMSE angular error (°): 19.82 -> 17.72 | Helped |
| Add Eikonal regularization (ablation) | PTv3 surface normal estimator · LiSu in-domain (20% test split) | RMSE angular error (°): 19.82 -> 17.74 | Helped |
| Add spatial GTV, temporal GTV, and Eikonal regularization (ablation) | PTv3 surface normal estimator · LiSu in-domain (20% test split) | RMSE angular error (°): 19.82 -> 17.58 | Helped |
| Use the full regularized self-trained normal estimator instead of LL1-only self-training (ablation) | ReSimAD neural surface reconstruction on Waymo · Waymo sequences (average) | RMSE: 2.00 -> 1.97 | Helped |
| Use the full regularized self-trained normal estimator instead of LL1-only self-training (ablation) | ReSimAD neural surface reconstruction on Waymo · Waymo sequences (average) | Chamfer Distance: 0.40 -> 0.37 | Helped |
Surprising findings
- Naive self-training without the regularization terms does not improve reconstruction over the baseline, whereas the regularized version does.
Limitations
- Evaluation of real-world normal accuracy is only qualitative because no real LiDAR datasets with normal annotations exist; real-world evaluation relies on indirect reconstruction proxies.
- Previous methods were evaluated on a downsampled test set (every fifth frame) due to their runtime cost.
Experimental setup
Datasets: LiSu Waymo Open Dataset PCPNet
Benchmarks: LiSu test split Waymo Open Dataset
Metrics: mean angular error median angular error RMSE accuracy at 5.0/7.5/11.25/22.5/30 degree thresholds Chamfer Distance runtime
Baselines: PCA Jet PCPNet CMG-Net GraphFit Du et al. NGL NeuralGF SHS-Net ReSimAD
Builds on: Point Transformer V3 CARLA ReSimAD PCPNet
Closest papers by meaning
- LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance (NeurIPS 2026 · similarity 0.87)
- High-quality Point Cloud Oriented Normal Estimation via Hybrid Angular and Euclidean Distance Encoding (CVPR 2025 · similarity 0.87)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.86)
- STGC-NeRF: Spatial-Temporal Geometric Consistency for LiDAR Neural Radiance Fields in Dynamic Scenes (AAAI 2025 · similarity 0.86)
- GeoNLF: Geometry guided Pose-Free Neural LiDAR Fields (NeurIPS 2024 · similarity 0.86)
- SS3DM: Benchmarking Street-View Surface Reconstruction with a Synthetic 3D Mesh Dataset (NeurIPS 2024 · similarity 0.86)
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.