Modern object detectors (e.g., YOLO, RT-DETR, and Faster R-CNN) can produce overconfident predictions for objects outside their training classes, which we refer to as out-of-distribution (OoD) objects, while the training classes are referred to as in-distribution (ID) classes. In practice, such predictions are semantically invalid: an OoD object is assigned a label from the object detector's ID classes with high confidence, even though the object falls outside the predefined label space. These failures can be viewed as instances of computer vision hallucination.</p>\n<p>SPK addresses a complementary limitation of existing state-of-the-art OoD detection methods, especially the post-hoc ones which typically make decisions in high-dimensional and opaque representation spaces produced by the target object detector. SPK aims to construct an interpretable and effective representation space for such methods. We find that this can be achieved by eliciting and structuring three complementary priors already encoded in the object detection model and its training data: a part-level semantic prior, indicating whether the predicted region supports the known class rather than proximal or background concepts; a geometric prior, capturing the plausibility of the object's relative scale; and a contextual prior, measuring whether the predicted class appears in a familiar visual environment. This strategy yields the low-dimensional, semantically grounded SPK space for OoD detection and mitigation.</p>\n","updatedAt":"2026-08-20T17:39:00.022Z","author":{"_id":"62f3896372a1894df39a60dd","avatarUrl":"/avatars/28f6b44a3bc7f05a16ebd3fe5ee7a9bc.svg","fullname":"WeichengHE","name":"HugoHE","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9083626866340637},"editors":["HugoHE"],"editorAvatarUrls":["/avatars/28f6b44a3bc7f05a16ebd3fe5ee7a9bc.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.19080","authors":[{"_id":"6a86ba40db13816030683fe9","name":"Changshun Wu","hidden":false},{"_id":"6a86ba40db13816030683fea","user":{"_id":"62f3896372a1894df39a60dd","avatarUrl":"/avatars/28f6b44a3bc7f05a16ebd3fe5ee7a9bc.svg","isPro":false,"fullname":"WeichengHE","user":"HugoHE","type":"user","name":"HugoHE"},"name":"Weicheng He","status":"claimed_verified","statusLastChangedAt":"2026-08-20T16:45:04.544Z","hidden":false},{"_id":"6a86ba40db13816030683feb","name":"Xiaowei Huang","hidden":false},{"_id":"6a86ba40db13816030683fec","name":"Saddek Bensalem","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/62f3896372a1894df39a60dd/EWlBufPErbjgrDXZvFd68.jpeg"],"publishedAt":"2026-08-19T00:00:00.000Z","submittedOnDailyAt":"2026-08-20T00:00:00.000Z","title":"SPK: Eliciting Structured Prior Knowledge for Interpretable Out-of-Distribution Detection in Real-Time Object Detection","submittedOnDailyBy":{"_id":"62f3896372a1894df39a60dd","avatarUrl":"/avatars/28f6b44a3bc7f05a16ebd3fe5ee7a9bc.svg","isPro":false,"fullname":"WeichengHE","user":"HugoHE","type":"user","name":"HugoHE"},"summary":"Object detectors often produce over-confident predictions for objects outside their training categories, leading to so-called out-of-distribution (OoD) hallucinations. Existing approaches for detecting or mitigating such hallucinations typically either construct scoring functions directly over learned object detector representations or modify the object detector itself to suppress hallucination emergence. However, the latent priors implicitly encoded in these representations remain largely unexplored and have not been explicitly decoded for OoD detection. To uncover and exploit these latent priors, we propose Structured Prior Knowledge (SPK), a hallucination-oriented framework that explicitly elicits OoD-relevant priors from pretrained object detectors. Specifically, SPK leverages in-distribution data and hallucination-inducing samples as diagnostic supervision to elicit part-level semantic concepts underlying object detector decision-making, rather than using them merely for rejection or object detector adaptation. The elicited semantic priors are further integrated with geometric and contextual priors to form a compact five-dimensional SPK representation for OoD detection. Extensive experiments across diverse object detector architectures and multiple OoD benchmarks demonstrate that SPK achieves state-of-the-art OoD detection. Our findings reveal that pretrained object detectors already encode substantially richer latent knowledge than is typically exploited for OoD detection. More importantly, this knowledge can be explicitly elicited and organized into a compact, structured, and interpretable knowledge space for prediction reliability analysis. This suggests a promising proactive route for improving object detector reliability by explicitly uncovering and leveraging latent priors. Code and data are available at: https://gricad-gitlab.univ-grenoble-alpes.fr/dnn-safety/spk","upvotes":0,"discussionId":"6a86ba40db13816030683fed","projectPage":"https://gricad-gitlab.univ-grenoble-alpes.fr/dnn-safety/spk","ai_summary":"Structured Prior Knowledge explicitly extracts latent semantic, geometric, and contextual priors from pretrained object detectors to detect out-of-distribution hallucinations via a compact interpretable representation.","ai_keywords":["out-of-distribution detection","object detectors","hallucinations","Structured Prior Knowledge","latent priors","part-level semantic concepts","geometric priors","contextual priors"],"ai_summary_model":"thinkingmachines/Inkling-Small","organization":{"_id":"656774a6ad0006b8524e90e7","name":"UoLiverpool","fullname":"University of Liverpool","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/656773ed05fb89fb7e14095d/JBKimcbjjRiaAbt0ytH2z.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[],"acceptLanguages":["en"],"organization":{"_id":"656774a6ad0006b8524e90e7","name":"UoLiverpool","fullname":"University of Liverpool","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/656773ed05fb89fb7e14095d/JBKimcbjjRiaAbt0ytH2z.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.19080.md","query":{}}">
SPK: Eliciting Structured Prior Knowledge for Interpretable Out-of-Distribution Detection in Real-Time Object Detection
Abstract
Structured Prior Knowledge explicitly extracts latent semantic, geometric, and contextual priors from pretrained object detectors to detect out-of-distribution hallucinations via a compact interpretable representation.
Object detectors often produce over-confident predictions for objects outside their training categories, leading to so-called out-of-distribution (OoD) hallucinations. Existing approaches for detecting or mitigating such hallucinations typically either construct scoring functions directly over learned object detector representations or modify the object detector itself to suppress hallucination emergence. However, the latent priors implicitly encoded in these representations remain largely unexplored and have not been explicitly decoded for OoD detection. To uncover and exploit these latent priors, we propose Structured Prior Knowledge (SPK), a hallucination-oriented framework that explicitly elicits OoD-relevant priors from pretrained object detectors. Specifically, SPK leverages in-distribution data and hallucination-inducing samples as diagnostic supervision to elicit part-level semantic concepts underlying object detector decision-making, rather than using them merely for rejection or object detector adaptation. The elicited semantic priors are further integrated with geometric and contextual priors to form a compact five-dimensional SPK representation for OoD detection. Extensive experiments across diverse object detector architectures and multiple OoD benchmarks demonstrate that SPK achieves state-of-the-art OoD detection. Our findings reveal that pretrained object detectors already encode substantially richer latent knowledge than is typically exploited for OoD detection. More importantly, this knowledge can be explicitly elicited and organized into a compact, structured, and interpretable knowledge space for prediction reliability analysis. This suggests a promising proactive route for improving object detector reliability by explicitly uncovering and leveraging latent priors. Code and data are available at: https://gricad-gitlab.univ-grenoble-alpes.fr/dnn-safety/spk
Community
Modern object detectors (e.g., YOLO, RT-DETR, and Faster R-CNN) can produce overconfident predictions for objects outside their training classes, which we refer to as out-of-distribution (OoD) objects, while the training classes are referred to as in-distribution (ID) classes. In practice, such predictions are semantically invalid: an OoD object is assigned a label from the object detector's ID classes with high confidence, even though the object falls outside the predefined label space. These failures can be viewed as instances of computer vision hallucination.
SPK addresses a complementary limitation of existing state-of-the-art OoD detection methods, especially the post-hoc ones which typically make decisions in high-dimensional and opaque representation spaces produced by the target object detector. SPK aims to construct an interpretable and effective representation space for such methods. We find that this can be achieved by eliciting and structuring three complementary priors already encoded in the object detection model and its training data: a part-level semantic prior, indicating whether the predicted region supports the known class rather than proximal or background concepts; a geometric prior, capturing the plausibility of the object's relative scale; and a contextual prior, measuring whether the predicted class appears in a familiar visual environment. This strategy yields the low-dimensional, semantically grounded SPK space for OoD detection and mitigation.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.19080 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.19080 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.19080 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.