LiON: Learning Point-wise Abstaining Penalty for LiDAR Outlier DetectioN Using Diverse Synthetic Data

TL;DR

LiON reframes LiDAR outlier detection as a selective classification problem and introduces a point-wise abstaining penalty learning framework to better separate inliers from outliers in semantically sparse point clouds. It combines a novel abstaining loss with a dynamic, point-wise calibration and a ShapeNet-based outlier synthesis pipeline that preserves LiDAR sampling patterns. The method achieves state-of-the-art results on SemanticKITTI and NuScenes across traditional and risk-coverage metrics, while maintaining efficient inference. This approach provides a practical, scalable pathway for open-world LiDAR perception in autonomous systems, with interpretable trade-offs between coverage and risk.

Abstract

LiDAR-based semantic scene understanding is an important module in the modern autonomous driving perception stack. However, identifying outlier points in a LiDAR point cloud is challenging as LiDAR point clouds lack semantically-rich information. While former SOTA methods adopt heuristic architectures, we revisit this problem from the perspective of Selective Classification, which introduces a selective function into the standard closed-set classification setup. Our solution is built upon the basic idea of abstaining from choosing any inlier categories but learns a point-wise abstaining penalty with a margin-based loss. Apart from learning paradigms, synthesizing outliers to approximate unlimited real outliers is also critical, so we propose a strong synthesis pipeline that generates outliers originated from various factors: object categories, sampling patterns and sizes. We demonstrate that learning different abstaining penalties, apart from point-wise penalty, for different types of (synthesized) outliers can further improve the performance. We benchmark our method on SemanticKITTI and nuScenes and achieve SOTA results. Codes are available at https://github.com/Daniellli/LiON/.

Figures (6)

• Figure 1: (a). The semantic segmentation model fails to identify furniture because the training set does not include such objects; (b). Comparison of our ShapeNet outlier synthesis method and the former resize outlier synthesis method.
• Figure 2: Method pipeline: a point cloud containing outliers synthesized by ShapeNet is processed by a feature extractor to obtain features. These features are then used by inlier and outlier classifiers to predict class logits.
• Figure 3: Our outlier synthesis pipeline: (a) loading a ShapeNet object; (b) randomly moving it away from the scene center; (c) randomly rotating it around the gravity direction; (d) randomly resizing it; (e) putting it on ground; (f) resampling points on the object to blend into the scene. We repeat this pipeline on the fly $G$ times for inserting $G$ objects.
• Figure 4: Statistics of outlier probability $\mathbf{p^o}$ for inlier and outlier samples under different settings of REAL on SemanticKITTI. For inliers, we would like to observe more samples with $\mathbf{p^o} \in [0,0.1]$, while for outliers, less samples with $\mathbf{p^o} \in [0,0.1]$ is desirable.
• Figure 5: (a). Comparison with the SOTA using the Risk-Coverage curves; (b). Comparison with different outlier synthesis pipeline using the risk-coverage curve; (c). Relationship between point-wise penalty loss and penalty $\alpha$.