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

Shaocong Xu, Pengfei Li, Qianpu Sun, Xinyu Liu, Yang Li, Shihui Guo, Zhen Wang, Bo Jiang, Rui Wang, Kehua Sheng, Bo Zhang, Li Jiang, Hao Zhao, Yilun ChenAAAI 2025 · OpenReview

Introduction

Autonomous vehicles rely heavily on LiDAR semantic segmentation to understand their surroundings. However, standard models are typically "closed-set," meaning they are trained to classify every point into a predefined set of categories (e.g., car, pedestrian, road). When these models encounter "out-of-distribution" (OOD) objects—such as furniture fallen from a truck or unusual road debris—they often "hallucinate," confidently misclassifying these outliers as known categories. This behavior poses a significant safety risk in autonomous driving pipelines.

The Challenge of Outlier Detection

Detecting outliers in LiDAR point clouds is uniquely difficult compared to 2D images because point clouds lack rich texture and color. Previous state-of-the-art (SOTA) methods attempted to synthesize outliers by resizing existing objects (e.g., the REAL method). However, this approach creates a "density shortcut": the model learns to identify outliers based on unnaturally high point densities resulting from shrinking objects, rather than learning actual shape or contextual cues.

Figure 1: (a) Standard segmentation models often misclassify out-of-distribution objects like furniture as road points. (b) Comparison between the previous REAL synthesis method and the proposed ShapeNet-based synthesis pipeline.

The LiON Framework

The authors introduce LiON (LiDAR Outlier DetectioN), a framework designed to overcome these limitations through two primary innovations:

1. Point-wise Abstaining Penalty

LiON reformulates outlier detection as Selective Classification (SC). In this setup, a model can choose to "abstain" from making a prediction if it is uncertain. Unlike traditional SC methods that use a global fixed penalty for abstaining, LiON learns a unique penalty ($\alpha$) for every single point. This point-wise approach is crucial for LiDAR data, where point density and classification difficulty vary significantly with distance from the sensor. By learning local penalties, the model better captures the geometric nuances of sparse versus dense regions.

2. Diverse ShapeNet Synthesis Pipeline

To provide the model with realistic training examples of outliers, the authors developed a synthesis pipeline using the ShapeNet dataset. Instead of simple resizing, they use a resampling technique that mimics the physics of a real LiDAR sensor. This ensures that synthetic outliers—ranging from chairs to tables—possess realistic point densities and occlusion patterns, preventing the model from relying on density shortcuts.

Experimental Evaluation

LiON was evaluated on major benchmarks, including SemanticKITTI and nuScenes, using ShapeNet as the source for unseen outlier objects. The framework significantly outperformed previous methods:

  • SemanticKITTI: Achieved an AUPR of 44.68 (an 8.58 point increase over the previous SOTA) and an AUROC of 92.69 (+7.09).
  • nuScenes: Achieved an AUPR of 31.58 (+10.38) and an AUROC of 95.24 (+10.74).

Qualitatively, LiON demonstrated a superior ability to identify "ego-cars" (the vehicle's own body) and unusual road objects that baseline models either missed or mislabeled as part of the drivable surface.

Conclusion

LiON represents a significant step forward in making LiDAR perception systems more robust to the unpredictable nature of real-world driving. By combining a physically grounded synthetic data pipeline with a point-wise selective classification strategy, it effectively mitigates the risks of model hallucination. Future research may focus on integrating these point-wise uncertainty scores directly into path-planning modules, allowing vehicles to respond more cautiously to high-penalty regions of the environment.

Summary updated

BibTeX
@inproceedings{xu2025lion,
  author = {Shaocong Xu and Pengfei Li and Qianpu Sun and Xinyu Liu and Yang Li and Shihui Guo and Zhen Wang and Bo Jiang and Rui Wang and Kehua Sheng and Bo Zhang and Li Jiang and Hao Zhao and Yilun Chen},
  title = {{LiON: Learning Point-Wise Abstaining Penalty for LiDAR Outlier DetectioN Using Diverse Synthetic Data}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=c6IQtImQNR},
}
Figure