SPK: Eliciting structured prior knowledge for interpretable out-of-distribution detection in real-time object detection

Researchers propose the structured prior knowledge framework SPK, which explicitly elicits latent priors from pretrained object detectors to achieve interpretable out-of-distribution (OoD) detection. The method uses in-distribution data and hallucination-inducing samples as diagnostic supervision to elicit part-level semantic concepts, and fuses them with geometric and contextual priors into a compact five-dimensional representation. Across multiple detector architectures and multiple OoD benchmarks, SPK achieves SOTA out-of-distribution detection performance, and the code and data have been made public.