SPK: Eliciting Structured Prior Knowledge for Interpretable Out-of-Distribution Detection in Real-Time Object Detection
SPK turns an object detector’s hidden priors into a five-dimensional signal for spotting hallucinated detections.
The paper targets over-confident detections on objects outside a model’s training categories. Its method uses in-distribution data and hallucination-inducing samples to elicit part-level semantic concepts from pretrained detectors, then combines those with geometric and contextual priors. The authors report state-of-the-art out-of-distribution detection across multiple detector architectures and benchmarks. HF Daily Papers' note
The paper targets over-confident detections on objects outside a model’s training categories. Its method uses in-distribution data and hallucination-inducing samples to elicit part-level semantic concepts from pretrained detectors, then combines those with geometric and contextual priors. The authors report state-of-the-art out-of-distribution detection across multiple detector architectures and benchmarks. HF Daily Papers' note
score 4