Research map / Point clouds / Segmentation and detection
LiON: Learning Point-Wise Abstaining Penalty for LiDAR Outlier DetectioN Using Diverse Synthetic Data
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.
Revisits LiDAR outlier detection through selective classification, proposing a point-wise abstaining penalty loss and a ShapeNet-based outlier synthesis pipeline that respects LiDAR sampling patterns, achieving SOTA on SemanticKITTI and nuScenes.
Problem
LiDAR outlier detection in driving scenes: point clouds lack semantically-rich information, so outlier points are hard to separate from inliers, and existing methods (REAL, APF) rely on calibration losses that hurt inlier accuracy and on resize-based synthesis that is unrealistic and limited in dive
Approach
Cast LiDAR outlier detection as selective classification with an extra outlier classifier, learn a point-wise abstaining penalty with margin-based losses in place of REAL's fixed calibration loss, and synthesize outliers by inserting ShapeNet objects with LiDAR-aware radius resampling.
Key results
- On SemanticKITTI, LiON reaches AUPR 44.68 and AUROC 92.69 versus REAL's 20.08 and 84.90 (C3D backbone).
- On nuScenes, LiON reaches AUPR 31.58 and AUROC 95.24, exceeding REAL's 21.20 and 84.50.
- Ablation on SemanticKITTI: removing the ShapeNet synthesis drops AUPR from 43.69 to 29.14, while removing resize synthesis drops it to 41.82.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Remove CCE loss from REAL (ablation) | REAL · SemanticKITTI | AUPR: 20.00 -> 26.68 | Helped |
| Remove CCE loss from REAL (ablation) | REAL · SemanticKITTI | AUROC: 84.90 -> 87.60 | Helped |
| Remove CCE loss from REAL (ablation) | REAL · SemanticKITTI | mIoU_old: 57.80 -> 58.28 | Helped |
| Add dynamic penalty instead of standard penalty (ablation) | LiON · SemanticKITTI | AUPR: 43.69 -> 44.68 | Helped |
| Add dynamic penalty instead of standard penalty (ablation) | LiON · SemanticKITTI | AUROC: 92.51 -> 92.69 | Helped |
| Add dynamic penalty instead of standard penalty (ablation) | LiON · SemanticKITTI | mIoU_old: 57.47 -> 57.56 | Helped |
| Add resize synthesis to ShapeNet synthesis (ablation) | LiON · SemanticKITTI | AUPR: 29.14 -> 43.69 | Helped |
| Add resize synthesis to ShapeNet synthesis (ablation) | LiON · SemanticKITTI | AUROC: 89.56 -> 92.51 | Helped |
| Add resize synthesis to ShapeNet synthesis (ablation) | LiON · SemanticKITTI | mIoU_old: 57.31 -> 57.47 | Helped |
| Add ShapeNet synthesis to resize synthesis (ablation) | LiON · SemanticKITTI | AUPR: 41.82 -> 43.69 | Helped |
| Add ShapeNet synthesis to resize synthesis (ablation) | LiON · SemanticKITTI | AUROC: 93.04 -> 92.51 | Hurt |
| Add ShapeNet synthesis to resize synthesis (ablation) | LiON · SemanticKITTI | mIoU_old: 57.30 -> 57.47 | Helped |
Surprising findings
- Removing REAL's calibration cross-entropy loss improves AUPR and AUROC on SemanticKITTI (20.00 to 26.68 AUPR), showing standard metrics can mislead about calibration effects.
- Resize-based synthesis alone gives a higher AUROC (93.04) than the combined pipeline (92.51) on SemanticKITTI, though the combined pipeline has the best AUPR.
Limitations
- Performance at low coverage (below ~90%) is worse than REAL, since the method rejects inliers as well as outliers at that point.
- Inlier mIoU drops slightly relative to closed-set Cylinder3D (57.56 vs 58.00 on SemanticKITTI) and is not uniformly improved on nuScenes baselines.
- Outlier synthesis relies on ShapeNet objects, which are not real LiDAR scans of outliers.
Experimental setup
Datasets: SemanticKITTI nuScenes ShapeNet
Benchmarks: SemanticKITTI nuScenes
Metrics: AUPR AUROC mIoU_old selective risk (mIoU-based) risk-coverage curves
Baselines: MSP MaxLogit MC-Dropout REAL APF
Builds on: REAL Cylinder3D learning to abstain outlier exposure
Closest papers by meaning
- Test-Time Training for LiDAR Semantic Segmentation under Corruption via Geometric Inlier Discrimination (CVPR 2026 · similarity 0.89)
- No Thing, Nothing: Highlighting Safety-Critical Classes for Robust LiDAR Semantic Segmentation in Adverse Weather (CVPR 2025 · similarity 0.87)
- PointCVaR: Risk-Optimized Outlier Removal for Robust 3D Point Cloud Classification (AAAI 2024 · similarity 0.86)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.86)
- Spotting the Unexpected (STU): A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving (CVPR 2025 · similarity 0.86)
- UniMix: Towards Domain Adaptive and Generalizable LiDAR Semantic Segmentation in Adverse Weather (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.