Research map / Detection and segmentation / 3D detection and driving
Is Your LiDAR Placement Optimized for 3D Scene Understanding?
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 Place3D, a full-cycle pipeline that scores multi-LiDAR placements with a semantic-occupancy entropy metric (M-SOG), optimizes placements with CMA-ES, and releases a 280k-frame CARLA dataset with clean and adverse conditions for LiDAR segmentation and 3D detection.
Problem
Multi-LiDAR placement on vehicles is chosen heuristically, and its effect on 3D scene understanding under adverse weather and sensor failures is unexplored; full data-collection-and-train cycles to test placements are too costly.
Approach
A surrogate metric (M-SOG) based on the entropy of probabilistic semantic occupancy grids, computed from ray-traversed voxels, predicts downstream perception quality without training; CMA-ES then searches LiDAR positions and roll angles to maximize it, with a theoretical optimality bound.
Key results
- Optimized placement 'Ours' achieves average adverse-condition mIoU of 48.6 (Mink) and 48.6 (SPV) for segmentation, vs best baseline ~47.1 (Pyramid SPV-style averages).
- For 3D detection, the optimized placement reaches average adverse mAP of 28.9 (Pillar), 29.5 (Center), and 26.1 (BEV), the highest among the placements shown.
- M-SOG scores correlate with clean-condition perception performance; the optimized placement has the highest M-SOG (-1.18e-6 detection, -1.29e-6 segmentation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the optimized Place3D placement instead of heuristic placements | LiDAR semantic segmentation · Place3D clean and adverse conditions | mIoU: up to 0.8% on clean datasets; up to 1.5% on corruption | Helped |
| Use the optimized Place3D placement instead of heuristic placements | 3D object detection · Place3D clean and adverse conditions | mAP: up to 0.7% on clean datasets; up to 0.3% on corruption | Helped |
| Optimize LiDAR placement using P-SOG derived from corrupted data rather than clean data (ablation) | Place3D · Adverse conditions | 3D perception performance | Helped |
| Constrain all LiDARs to the same height and optimize their horizontal-plane positions (ablation) | Place3D 2D-plane placement · Clean and corruption settings | 3D perception performance | Helped |
| Add roll-angle adjustments to Pyramid placement (ablation) | 3D semantic segmentation · Place3D | mIoU | Helped |
| Add roll-angle adjustments to Line placement (ablation) | 3D semantic segmentation · Place3D | mIoU | Hurt |
| Add roll-angle adjustments to Pyramid and Line placements (ablation) | 3D object detection · Place3D | mAP | Mixed |
| Increase the average height of the four LiDARs | LiDAR placement · Place3D | 3D perception performance | Helped |
| Increase the difference among LiDAR heights | LiDAR placement · Place3D | 3D perception performance | Helped |
| Use a more uniform, spread-out four-LiDAR distribution | LiDAR placement · Place3D semantic segmentation | mIoU | Helped |
Surprising findings
- Under corruption, the top-performing clean-condition baselines can be notably worse, and the M-SOG correlation with performance is weaker under adverse conditions.
Limitations
- Uses low-resolution 16-channel LiDARs in simulation only, so sim-to-real transfer is not validated.
- Only four LiDARs and a small set of heuristic baselines; M-SOG correlation with performance is weaker under corruption.
Experimental setup
Datasets: Place3D (CARLA-generated) nuScenes split ratio KITTI class configuration
Benchmarks: Place3D
Metrics: mIoU mAP M-SOG
Baselines: Center Line Pyramid Square Trapezoid Line-roll Pyramid-roll S-MIG
Builds on: S-MIG CMA-ES Robo3D
Closest papers by meaning
- Investigating the Impact of Multi-LiDAR Placement on Object Detection for Autonomous Driving (CVPR 2022 · similarity 0.92)
- ALSO: Automotive Lidar Self-Supervision by Occupancy Estimation (CVPR 2023 · similarity 0.88)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.87)
- TREND: Unsupervised 3D Representation Learning via Temporal Forecasting for LiDAR Perception (NeurIPS 2025 · similarity 0.87)
- OLiDM: Object-aware LiDAR Diffusion Models for Autonomous Driving (AAAI 2025 · similarity 0.87)
- HiLoTs: High-Low Temporal Sensitive Representation Learning for Semi-Supervised LiDAR Segmentation in Autonomous Driving (CVPR 2025 · 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.