Exploring Implicit Conceptual Patterns for Out-of-Distribution Object Detection

Luyao Tang, Kunze Huang, Zheyuan Cai

Abstract: Reliable out-of-distribution (OOD) object identification is critical for robust deployment of detectors in real-world scenarios. Existing methods typically rely on instance-level feature adjustment or pseudo-anomaly generation to endow models with unknown-aware capabilities, but they struggle to capture the intrinsic nature of "unknown" due to the lack of authentic OOD samples. This paper presents Concept-Aware Contextual Reasoner, a concise yet powerful framework designed for OOD object detection. Our core insight is that ID and OOD objects often coexist in complex scenes, requiring contextual semantic reasoning to distinguish their inherent differences through implicit concept modeling. The framework includes two key components: unsupervised conceptual mining and contextual concept fusion. The former adopts a scene-aware learning paradigm to abstract holistic image information—covering both ID and OOD objects—and generates compact slot-like semantic representations with relational constraints. The latter dynamically integrates these semantic slots with object candidates from the base detector, anchoring the "unknown" concept into the detection pipeline. During inference, a scene-aware OOD confidence scoring mechanism is introduced to enhance the discriminability between ID and OOD objects. Extensive experiments on standard benchmarks verify the framework’s effectiveness, reducing FPR95 by up to 13.2% compared with the state-of-the-art methods.