Object Co-Occurrence Framework Improves OOD Detection

A novel framework for out-of-distribution (OOD) detection, centered on objects, utilizes co-occurrence patterns to mitigate simplicity bias present in deep learning models. Introduced in arXiv:2605.07821, this technique captures Object CO-occurrence (OCO) patterns by predicting disentangled representations for test samples. It further categorizes these patterns into three distinct scenarios to enhance the differentiation between in-distribution and near-OOD data. Drawing inspiration from how the human visual system interprets object co-occurrence for scene comprehension, this framework tackles the issue of identifying near-OOD instances, where simplicity bias in models hampers the learning of discriminative features in disentangled representations. Current approaches primarily emphasize regular entangled representations, overlooking the rich contextual details within images.