Mitigating Simplicity Bias in OOD Detection through Object Co-occurrence Analysis

The article presents a novel Object-Centric OOD detection framework that leverages Object CO-occurrence (OCO) patterns to improve the detection of out-of-distribution (OOD) samples, particularly near-OOD instances. The method predicts disentangled representations for test samples and categorizes them based on observed co-occurrence patterns in the training data, allowing for a more nuanced detection process that incorporates semantic relationships. Experimental evaluations show that OCO achieves competitive results in various OOD scenarios, addressing both semantic and covariate shifts, with the code available at https://github.com/Michael-McQueen/OCO, making it a valuable tool for practitioners in enhancing model reliability.