SeNeDiF-OOD: Semantic Nested Dichotomy Fusion for Out-of-Distribution Detection Methodology in Open-World Classification. A Case Study on Monument Style Classification

Out-of-distribution (OOD) detection is a fundamental requirement for the reliable deployment of artificial intelligence applications in open-world environments. However, addressing the heterogeneous nature of OOD data, ranging from low-level corruption to semantic shifts, remains a complex challenge that single-stage detectors often fail to resolve. To address this issue, we propose SeNeDiF-OOD, a novel methodology based on Semantic Nested Dichotomy Fusion. This framework decomposes the detection task into a hierarchical structure of binary fusion nodes, where each layer is designed to integrate decision boundaries aligned with specific levels of semantic abstraction. To validate the proposed framework, we present a comprehensive case study using MonuMAI, a real-world architectural style recognition system exposed to an open environment. This application faces a diverse range of inputs, including non-monument images, unknown architectural styles, and adversarial attacks, making it an ideal testbed for our proposal. Through extensive experimental evaluation in this domain, results demonstrate that our hierarchical fusion methodology significantly outperforms traditional baselines, effectively filtering these diverse OOD categories while preserving in-distribution performance.

💡 Research Summary

The paper introduces SeNeDiF‑OOD (Semantic Nested Dichotomy Fusion for Out‑of‑Distribution detection), a hierarchical framework that reframes OOD detection as a series of semantically grounded binary decisions. Rather than relying on a single global rejection score, the method decomposes the input space into successive checkpoints that correspond to increasing levels of semantic abstraction. Each checkpoint is implemented as a binary node in a nested‑dichotomy tree; the tree’s topology is aligned with human‑defined semantic hierarchies (e.g., “is the image a monument?”, “does it belong to a supported architectural style?”, “is the image quality sufficient?”, etc.). At every node, the authors fuse the most appropriate OOD detection mechanism—softmax confidence, Mahalanobis distance, energy‑based scores, reconstruction error, or a combination thereof—tailoring the decision rule to the specific data regime of that level. The binary outcomes are probabilistically fused, yielding a final OOD score that reflects both learned model evidence and prior semantic knowledge.

The theoretical contribution includes a formal analysis showing that hierarchical binary classification reduces the learning complexity from O(k) to O(log k) for k classes, and that error propagation is limited because each layer can reject samples independently. When the dichotomy tree mirrors a meaningful semantic hierarchy, performance gains stem more from this alignment than from traditional tree‑selection strategies (random, balanced, performance‑based).

To validate the approach, the authors embed SeNeDiF‑OOD into MonuMAI, a real‑world mobile application that classifies architectural styles (four styles: Hispanic‑Muslim, Gothic, Renaissance, Baroque). Over four years of operation, MonuMAI receives a wide variety of out‑of‑distribution inputs: non‑monument photos, low‑quality captures, buildings of unsupported styles, and deliberately crafted adversarial images. The authors collected ~4,000 user‑uploaded images and supplemented them with public OOD datasets, labeling them into five categories: in‑distribution, non‑monument, low‑quality, unsupported style, and adversarial attack.

Experimental results demonstrate that SeNeDiF‑OOD dramatically reduces false positive detections. Compared with the original MonuMAI model (which uses a single softmax‑based OOD score), the hierarchical system lowers the overall OOD false‑positive rate from 27 % to under 8 %. At the top level, non‑monument images are filtered with >95 % accuracy; the middle layers successfully reject low‑quality images (~90 % recall) and unsupported styles (~92 % precision). Adversarial attacks are identified with ~89 % precision. Importantly, in‑distribution classification accuracy remains essentially unchanged (93.4 % vs. 92.7 % for the baseline), confirming that the added rejection mechanisms do not harm normal performance.

Layer‑wise analysis reveals that early checkpoints capture coarse OOD types (e.g., “not a monument”), while deeper nodes discriminate finer semantic shifts (e.g., “unknown architectural style”). This structure provides interpretable diagnostics: when a sample is rejected at a particular node, the system can report the specific reason, facilitating monitoring, data curation, and targeted model updates. Moreover, the final node can serve as an active‑learning trigger, flagging samples that may define new classes for future training.

The authors argue that the methodology is broadly applicable beyond architectural style classification. Any domain with a natural semantic hierarchy—medical imaging (normal vs. pathological vs. novel disease), autonomous driving (road scene vs. off‑road vs. unexpected object), or multimodal conversational agents—could benefit from a similar nested‑dichotomy OOD framework. Future work will explore automatic hierarchy learning, extension to multimodal inputs, and computational optimizations for real‑time deployment.