Out-of-Distribution Detection in LiDAR Semantic Segmentation Using Epistemic Uncertainty from Hierarchical GMMs

TL;DR

This paper tackles unsupervised out-of-distribution detection in LiDAR semantic segmentation by isolating epistemic uncertainty. It models per-class feature distributions with Gaussian Mixture Models and places hierarchical Bayesian priors over GMM parameters to quantify uncertainty. Inference samples GMM parameters, uses majority voting for semantic predictions, and flags high epistemic uncertainty pixels as OOD without needing OOD data or retraining. On SemanticKITTI, the approach yields state-of-the-art AUROC and AUPRC and lower FPR95 while preserving segmentation accuracy, demonstrating practical potential for safer autonomous perception.

Abstract

In addition to accurate scene understanding through precise semantic segmentation of LiDAR point clouds, detecting out-of-distribution (OOD) objects, instances not encountered during training, is essential to prevent the incorrect assignment of unknown objects to known classes. While supervised OOD detection methods depend on auxiliary OOD datasets, unsupervised methods avoid this requirement but typically rely on predictive entropy, the entropy of the predictive distribution obtained by averaging over an ensemble or multiple posterior weight samples. However, these methods often conflate epistemic (model) and aleatoric (data) uncertainties, misclassifying ambiguous in distribution regions as OOD. To address this issue, we present an unsupervised OOD detection approach that employs epistemic uncertainty derived from hierarchical Bayesian modeling of Gaussian Mixture Model (GMM) parameters in the feature space of a deep neural network. Without requiring auxiliary data or additional training stages, our approach outperforms existing uncertainty-based methods on the SemanticKITTI dataset, achieving an 18\% improvement in AUROC, 22\% increase in AUPRC, and 36\% reduction in FPR95 (from 76\% to 40\%), compared to the predictive entropy approach used in prior works.

Figures (3)

• Figure 1: Proposed epistemic uncertainty-based OOD detection. A deep neural network extracts a 32-dimensional feature space from 5D LiDAR input. Class-conditional GMMs model the feature distribution per class. During inference, epistemic uncertainty is quantified by aggregating multiple predictions obtained through sampling GMM parameters (mean and variance) from their posterior distributions. Pixels exhibiting higher epistemic uncertainty scores are considered more likely to represent OOD instances.
• Figure 2: Comparison of predictive entropy and epistemic uncertainty for OOD detection. The proposed epistemic uncertainty measure offers a more precise separation between ID and OOD samples.
• Figure 3: Qualitative comparison of OOD detection using epistemic uncertainty from our proposed approach versus predictive entropy from DE across four representative LiDAR scenes. From top to bottom, the rows depict: (1) Semantic predictions with class color coding: outlier, car, road, sidewalk, building, fence, vegetation, trunk, terrain, pole; (2) OOD ground truth, where ID pixels are black and OOD pixels are yellow; (3) Epistemic uncertainty map from our method and (5) predictive entropy map from DE, visualized with a temperature scale from low to high uncertainty; (4) OOD detection results from our method and (7) DE, highlighting detected OOD regions in yellow; (5) Error map highlighting OOD objects not identified by our method, marked in red; (8) Camera image for visual reference.