Relative Energy Learning for LiDAR Out-of-Distribution Detection
Researchers have developed a new framework called Relative Energy Learning (REL) for detecting out-of-distribution (OOD) objects in 3D LiDAR point clouds, a crucial task for autonomous driving safety. Unlike previous methods that struggled with distinguishing anomalies, REL uses the energy gap between in-distribution and out-of-distribution logits to improve robustness. To overcome the lack of OOD data during training, the team introduced a lightweight data synthesis strategy called Point Raise, which generates auxiliary anomalies by perturbing existing point clouds. Experiments on SemanticKITTI and the STU benchmark showed REL significantly outperforms existing methods. AI
IMPACT Enhances safety for autonomous vehicles by improving the detection of unexpected objects in real-world driving scenarios.