OCCUQ: Exploring Efficient Uncertainty Quantification for 3D Occupancy Prediction

TL;DR

This paper tackles robust 3D occupancy prediction for autonomous driving by introducing an efficient Uncertainty Quantification Module that attaches to a multi-camera SurroundOCC backbone. It leverages a Gaussian Mixture Model over intermediate features to estimate epistemic uncertainty and employs an uncertainty-guided calibration strategy to maintain reliable confidence under distribution shifts and corruptions. The approach demonstrates improved OoD detection and calibration at both scene and region levels, with minimal computational overhead compared to Deep Ensembles or MC-Dropout. The work advances practical deployment by enabling reliable uncertainty estimates in real-time perception, and provides code and data access for reproducibility and further research.

Abstract

Autonomous driving has the potential to significantly enhance productivity and provide numerous societal benefits. Ensuring robustness in these safety-critical systems is essential, particularly when vehicles must navigate adverse weather conditions and sensor corruptions that may not have been encountered during training. Current methods often overlook uncertainties arising from adversarial conditions or distributional shifts, limiting their real-world applicability. We propose an efficient adaptation of an uncertainty estimation technique for 3D occupancy prediction. Our method dynamically calibrates model confidence using epistemic uncertainty estimates. Our evaluation under various camera corruption scenarios, such as fog or missing cameras, demonstrates that our approach effectively quantifies epistemic uncertainty by assigning higher uncertainty values to unseen data. We introduce region-specific corruptions to simulate defects affecting only a single camera and validate our findings through both scene-level and region-level assessments. Our results show superior performance in Out-of-Distribution (OoD) detection and confidence calibration compared to common baselines such as Deep Ensembles and MC-Dropout. Our approach consistently demonstrates reliable uncertainty measures, indicating its potential for enhancing the robustness of autonomous driving systems in real-world scenarios. Code and dataset are available at https://github.com/ika-rwth-aachen/OCCUQ .

Figures (4)

• Figure 1: Our method introduces a lightweight Uncertainty Quantification Module to the main model, that provides uncertainties for 3D occupancy fields. Motorcyclists in the distance has higher uncertainty scores indicating it as a rare class in the training set.
• Figure 2: Overview of the proposed method. From multi-view camera images our method provides 3D occupancy predictions with reliable epistemic and aleatoric uncertainties at voxel level. We build on top of the SurroundOCC wei2023surroundocc model, and introduce an uncertainty quantification (UQ) module.
• Figure 3: Feature density under data corruption. The corruption type Snow significantly impacts the feature density distribution of our proposed method leading to a near-complete separation between ID and OoD samples at severity level 3. In contrast, baseline approaches exhibit a less pronounced response to this shift in input distribution.
• Figure 4: Construction site scenario. The front camera (\ref{['fig:sample952_front']}) captures a challenging scene with numerous traffic cones, rare traffic sign objects, nearby a construction site. While baseline models (\ref{['fig:sample952_se']}), (\ref{['fig:sample952_de5pe']}) and (\ref{['fig:sample952_mcd5_pe']}) fail to present the uncertainty in this scenario, the feature density (\ref{['fig:sample952_feature_density']}) reveals high uncertainty near and around these objects, including the construction site.