Title: Relative Energy Learning for LiDAR Out-of-Distribution Detection

Abstract:

Ensuring safety in autonomous driving requires robust out-of-distribution (OOD) detection, as vehicles must reliably identify road hazards and unfamiliar objects that fall outside their training data. While OOD detection has been widely studied in 2D imagery, directly applying these techniques to 3D LiDAR point clouds has proven unsuccessful. Existing LiDAR-based OOD methods often fail to differentiate between rare anomalies and standard classes, resulting in significant false positives and overconfident mistakes in critical safety scenarios.

To address these challenges, we introduce Relative Energy Learning (REL), a straightforward yet powerful framework designed for OOD detection in LiDAR point clouds. REL utilizes the energy disparity between positive (in-distribution) and negative logits as a relative scoring mechanism. This approach alleviates calibration problems associated with raw energy values and enhances model robustness across diverse environments. Furthermore, to overcome the lack of OOD samples during the training phase, we propose a lightweight data synthesis technique named Point Raise. This method generates auxiliary anomalies by perturbing existing point clouds without modifying the underlying semantics of inliers.

Benchmark evaluations on SemanticKITTI and the Spotting the Unexpected (STU) dataset demonstrate that REL significantly surpasses current state-of-the-art methods. Our findings indicate that modeling relative energy, paired with simple synthetic outliers, offers a principled and scalable approach to achieving reliable OOD detection for autonomous driving in open-world conditions.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC