Uncertainty Estimation for 3D Semantic Segmentation: MoCA: Uncertainty Estimation through Monte Carlo Augmentations for 3D Unsupervised Domain Adaptation

Reliable uncertainty estimation is crucial for deploying machine learning models in safety critical domains such as robotics and autonomous driving, where incorrect predictions can lead to severe consequences. Uncertainty in machine learning is typically categorized into aleatoric and epistemic components, with epistemic uncertainty being particularly relevant for models processing 3D point cloud data, because it captures the model’s lack of knowledge. These models are especially vulnerable to domain shifts caused by variations in sensing technologies and reconstruction pipelines. Existing approaches to uncertainty estimation in 3D models often require architectural changes or computationally expensive retraining, limiting their practicality. To address this, this thesis introduces MoCA, a plug-and-play, backbone-agnostic method for uncertainty estimation in 3D point cloud neural networks. The method is based on the observation that models respond less consistently to strong augmentations when faced with out-of-distribution inputs. Experimental results show that MoCA provides more reliable uncertainty estimates than existing techniques and effectively distinguishes between correct and incorrect predictions. Furthermore, all methods are evaluated on the downstream task of source-free unsupervised domain adaptation for semantic segmentation, where MoCA achieves significant improvements over prior approaches.