DOSE3 : Diffusion-Based Out-of-Distribution Detection on SE(3) Trajectories

Abstract

OOD detection is a machine learning task that seeks to identify abnormal samples. Traditionally, OOD detection requires model retraining for different inliner distributions. Recent work has shown that diffusion models can be applied to OOD detection tasks, but these attempts have either assumed that the samples are distributed in Euclidean space or a latent image space. In this work, we advance OOD to trajectories in the Special Euclidean Group in 3D ($\mathbb{SE}(3)$). In particular, many tasks in computer vision, robotics and engineering disciplines require reasoning about data in the form of sequences of poses of objects that are in the $\mathbb{SE}(3)$ To this end, we introduce the $\textbf{D}$iffusion-based $\textbf{O}$ut-of-distribution detection on $\mathbb{SE}(3)$ framework, $\mathbf{DOSE3}$, a novel OOD framework which extends diffusion to a unified sample space of $\mathbb{SE}(3)$ pose sequences. We validate our approach on OOD detection tasks on multiple benchmark datasets, and demonstrate the efficacy of our approach against state-of-the-art OOD detection frameworks.