The Good, the Bad, and the Template: Contrastive Anomaly Detection in 3D

Anomaly detection and localization (ADL) in point clouds is a rapidly expanding field of 3D computer vision, owing to its importance in robotic manufacturing and automated quality control. The most recent ADL methods focus on extracting representations of 3D geometries from a test object and comparing these to representations of anomaly-free objects. Despite rapid progress, modern ADL techniques still struggle to meet accuracy expectations in safety-critical fields such as aerospace. The main challenges are learning representations that are highly robust for anomaly detection, and developing algorithms that accurately and unambiguously compare these representations. We overcome the first challenge by formulating ADL as a contrastive learning problem and develop a deep representation extractor, highly optimized for anomaly detection. For the second challenge, we compare test representations with anomaly-free representations of multiple reference objects, precisely aligned in the common 3D reference frame. Our method establishes the new state of the art on both Real3D-AD and Anomaly-Shapenet datasets, reaching mean area under ROC curve of 0.91 and 0.95 respectively.

Conference paper