Abstract
This paper introduces a novel anomaly detection (AD) problem that focuses onidentifying `odd-looking' objects relative to the other instances within ascene. Unlike the traditional AD benchmarks, in our setting, anomalies in thiscontext are scene-specific, defined by the regular instances that make up themajority. Since object instances are often partly visible from a singleviewpoint, our setting provides multiple views of each scene as input. Toprovide a testbed for future research in this task, we introduce twobenchmarks, ToysAD-8K and PartsAD-15K. We propose a novel method that generates3D object-centric representations for each instance and detects the anomalousones through a cross-examination between the instances. We rigorously analyzeour method quantitatively and qualitatively in the presented benchmarks.