A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges

TL;DR

This paper provides a comprehensive cross-domain survey of anomaly, novelty, open-set, and out-of-distribution detection, unifying these related tasks under a shared taxonomy and highlighting their interconnections. It catalogs a wide range of methods—from classical unsupervised detectors to deep generative, self-supervised, and hybrid discriminative-generative approaches—within a single hierarchical framework. The survey also catalogs datasets and evaluation protocols, analyzes core challenges, and outlines practical and theoretical directions for fairness, robustness, explainability, and open-world recognition. By exposing cross-domain synergies and common principles (e.g., compact normal representations, outlier exposure, prototype-based reasoning, and SSL signals), the work aims to spur unified methodological advances and more realistic evaluations.

Abstract

Machine learning models often encounter samples that are diverged from the training distribution. Failure to recognize an out-of-distribution (OOD) sample, and consequently assign that sample to an in-class label significantly compromises the reliability of a model. The problem has gained significant attention due to its importance for safety deploying models in open-world settings. Detecting OOD samples is challenging due to the intractability of modeling all possible unknown distributions. To date, several research domains tackle the problem of detecting unfamiliar samples, including anomaly detection, novelty detection, one-class learning, open set recognition, and out-of-distribution detection. Despite having similar and shared concepts, out-of-distribution, open-set, and anomaly detection have been investigated independently. Accordingly, these research avenues have not cross-pollinated, creating research barriers. While some surveys intend to provide an overview of these approaches, they seem to only focus on a specific domain without examining the relationship between different domains. This survey aims to provide a cross-domain and comprehensive review of numerous eminent works in respective areas while identifying their commonalities. Researchers can benefit from the overview of research advances in different fields and develop future methodology synergistically. Furthermore, to the best of our knowledge, while there are surveys in anomaly detection or one-class learning, there is no comprehensive or up-to-date survey on out-of-distribution detection, which our survey covers extensively. Finally, having a unified cross-domain perspective, we discuss and shed light on future lines of research, intending to bring these fields closer together.

Figures (43)

• Figure 1: Problem setup for ND, OSR, and ODD from a unified perspective based on the common routine followed in the respective fields. (A), (B) and, (C) are sampled from the same training dataset, while (D) is sampled from different datasets. Typically, in ND, all training samples are deemed normal, and share a common semantic (green region), while samples diverged from such distribution are considered as anomalous. Although samples in area (D) can be regarded as potential outliers, only areas (B) and (C) are used as anomalies for the evaluation phase. In OSR, more supervision is available by accessing the label of normal samples. For example "car", "dog", "cat", "airplane" and, "bus" classes i.e, the union of (A) and (B), are considered as normal while (C) is open-set distribution(see right Figure). Same as ND, (D) is not usually considered an open-set distribution in the field, while there is no specific constraint on the type of open-set distribution in the definition of OSR domain. In OOD detection, multiple classes are considered as normal, which is quite similar to OSR. For example, (A), (B), and (C) comprise the normal training distributions, and another distribution that shows a degree of change with respect to the training distribution and is said to be out-of-distribution, which can be (D) in this case.
• Figure 2: As discussed in section \ref{['sec.2']}, all explained approaches can be unitedly classified in the shown hierarchical structure. The tree on the right points out that although some approaches have been labeled as ND, OSR, or OOD detection in the field, however can be classified in a more grounded and general form such that their knowledge could be shared synergistically. For instance, self-supervised ND methods can be added to multi-class classification approaches without harming their classification assumptions. Unfamiliar sample detection can be done by employing different levels of supervision, which is shown on the left.
• Figure 3: An overview of the isolation forest method is shown. Anomalies are more susceptible to isolation and hence have short path lengths. A normal point $x_i$ requires twelve random partitions to be isolated compared to an anomaly $x_o$ that requires only four partitions to be isolated. The figure is taken from liu2008isolation.
• Figure 4: The minimal number of points required to generate a dense zone in this figure is four. Since the area surrounding these points in a radius contains at least four points, Point A and the other red points are core points (including the point itself). They constitute a single cluster because they are all reachable from one another. B and C are not core points, but they may be reached from A (through other core points) and hence are included in the cluster. Point N is a noise point that is neither a core nor reachable directly. The figure is taken from ester1996density.
• Figure 5: An overview of the ALOCC method is shown. This work trains an autoencoder that fools a discriminator in a GAN-based setting. This helps the AE to make high-quality images instead of blurred outputs. Besides, by using the discriminator's output, a more semantic similarity loss is employed instead of the pixel-level $L_2$ reconstruction loss. The figure is taken from sabokrou2018adversarially.