Unknown-Aware Out-of-Distribution Object Detection via Unsupervised Concept Slots and Neural Binding

Luyao Tang, Kunze Huang, Yuxuan Yuan

Abstract: Detecting out-of-distribution (OOD) objects is critical for safely deploying object detectors in the wild. While prior work encourages unknown-aware behavior by regularizing instance-level feature spaces (e.g., via outlier synthesis), internalizing the notion of “unknown” without real unknown data remains challenging. We present COS-Binder, a lightweight yet effective framework for OOD object detection. Our central insight is that in object detection, in-distribution (ID) and OOD objects frequently co-occur within the same scene; thus, global abstraction and reasoning are needed to disentangle them. COS-Binder has two components: (i) an unsupervised concept discovery module that adopts an object-centric paradigm to summarize images into sparse, compressed, slot-based representations under relational constraints, and (ii) a neural concept binder that dynamically couples these slots with detector proposals, injecting unknown-aware cues into the base detector. At inference, we compute an image-guided OOD score that exploits slot–proposal consistency to reinforce the separation. Experiments on standard benchmarks demonstrate consistent gains.