Generalized ODIN: Detecting Out-of-distribution Image without Learning from Out-of-distribution Data

Deep learning models have achieved remarkable success in controlled environments, but their robustness in open-world settings remains limited. Early methods like OpenMax, DeepEnsemble, and temperature scaling improved confidence calibration but still relied heavily on labeled OoD data for tuning. Recent approaches such as ODIN and Mahalanobis distance-based detectors introduced input perturbation and feature space analysis, respectively, significantly enhancing detection performance. However, these methods depend on hyperparameters tuned with OoD samples, which limits their practicality. As datasets grow larger and more complex, the challenge of detecting diverse and unseen distribution shifts—semantic (new classes) and non-semantic (appearance variations)—becomes more pressing. The need for a method that can operate effectively without prior OoD data has become a key research focus, especially for deployment in safety-critical applications like autonomous driving, medical diagnosis, and security systems.

The core challenge is to develop OoD detection techniques that do not require prior access to OoD data for tuning hyperparameters. Existing methods often suffer from poor generalization when faced with unseen distribution shifts or novel classes, leading to false positives or missed detections. This is particularly problematic in real-world scenarios where collecting comprehensive OoD datasets is infeasible. The fundamental question is how to design models that can reliably distinguish in-distribution from out-of-distribution samples without relying on explicit OoD examples, thereby ensuring robustness, scalability, and ease of deployment across diverse applications.

The paper introduces a probabilistic confidence decomposition framework, which models the classifier's output as a ratio of joint class-domain probability and domain probability, explicitly incorporating a binary variable d indicating in/out distribution. The classifier architecture is modified to include hi(x) and g(x) branches, representing class confidence and shift degree, respectively. This structure allows the model to learn from in-distribution data alone, using input perturbation tuned via validation data without OoD samples. The approach generalizes ODIN by replacing hyperparameter tuning with an adaptive, data-driven mechanism, leading to superior detection performance and robustness across datasets and shift types.

• �� Define the joint class-domain probability p(y, din|x) and the domain variable d.
• �� Design a classifier with two branches: hi(x) modeling class confidence, g(x) modeling shift degree.
• �� Use softmax normalization of hi(x) and g(x) to produce class probabilities conditioned on d.
• �� Implement input perturbation by searching for an optimal magnitude only on in-distribution validation data, avoiding OoD samples.
• �� Train the model by minimizing cross-entropy loss on p(y, din|x), encouraging hi(x) and g(x) to behave as factors of confidence and shift.
• �� Evaluate detection scores based on hi(x) or g(x), with thresholds for rejection.
• �� Explore different similarity measures (inner product, cosine, Euclidean) for hi(x) to enhance robustness.
• �� Incorporate temperature scaling via g(x) for calibration and improved separation of OoD samples.

The evaluation employs datasets like CIFAR-10/100, TinyImageNet, LSUN, SVHN, and DomainNet, comparing the proposed method with ODIN and Mahalanobis baselines. The experiments focus on scenarios without OoD data for tuning, using validation sets to automatically select perturbation magnitudes. Performance metrics include AUROC and TNR@TPR95, assessed across multiple shift types. Ablation studies analyze the impact of different classifier structures, similarity measures, and perturbation strategies. Results consistently show that the proposed confidence decomposition and input tuning outperform traditional methods, with AUROC improvements of over 7% in some cases, and robustness across datasets and shift types.

On CIFAR-100, the method achieves AUROC scores exceeding 97%, with TNR@TPR95 over 87%, outperforming ODIN (AUROC 90.5%) and Mahalanobis (AUROC 92.4%). In large-scale DomainNet, detection of semantic shifts is easier than non-semantic shifts, but the method maintains high performance in both cases. The no-OoD-tuning version performs comparably to tuned SOTA methods, demonstrating excellent generalization. The analysis of different shift types reveals that non-semantic shifts pose greater detection challenges, guiding future research to focus on feature robustness and shift modeling.

This approach is suitable for deployment in autonomous vehicles, security systems, and medical diagnostics, where collecting comprehensive OoD data is impractical. It enables reliable detection of unknown inputs with minimal supervision, reducing false alarms and enhancing safety. The method's adaptability to various datasets and shift types makes it a versatile tool for real-world AI applications. Future integration with self-supervised learning and multimodal data could further improve detection accuracy and robustness in complex environments.

While effective, the method's performance may decline under extreme or highly novel distribution shifts, especially when the assumptions about shift factors do not hold. Computational complexity increases with high-dimensional data and complex models, limiting real-time application. Although tuning only on in-distribution data simplifies deployment, its effectiveness across diverse tasks and datasets requires further validation. Future work should address these limitations by developing more scalable algorithms and theoretical guarantees for detection under various shift scenarios.