Concept-Aware Representation Learning for Robust Out-of-Distribution Object Detection

Luyao Tang, Kunze Huang, Jiewei Zheng

Abstract: Modern object detectors often fail catastrophically when encountering novel objects outside their training distribution—a critical weakness for real-world applications. Traditional solutions rely on instance-level feature regularization or synthetic outlier generation, yet these approaches fundamentally lack genuine understanding of what constitutes "unknown." We hypothesize that this limitation stems from treating OOD detection as a localized classification problem rather than a holistic scene understanding task. To address this, we develop ConceptOD, which learns to represent images through decomposed semantic concepts without supervision. At its core, ConceptOD employs slot-based attention mechanisms to discover and encode latent visual concepts from unlabeled data, creating a structured representation space where both familiar and unfamiliar objects can be systematically characterized. These learned concept representations are then integrated with standard detection pipelines via a trainable binding mechanism, enabling detectors to leverage global scene semantics when evaluating individual proposals. Our scoring strategy further exploits the relationship between local object features and global concept distributions to identify OOD instances. Evaluated across multiple benchmarks, ConceptOD consistently outperforms existing approaches. These results suggest that unsupervised concept learning provides a principled pathway toward more reliable open-world object detection.