Test-Time Training for Out-of-Distribution Industrial Anomaly Detection via Robust Distribution Alignment
Keywords: Anomaly Detection, Test Time Training; Out-of-Distribution
Abstract: Detecting anomalous patterns is essential for quality control in industrial applications, with state-of-the-art methods relying on large defect-free datasets to model normal distributions. However, robustness under domain shift, such as changes in lighting or sensor drift, remains a critical challenge in real-world deployment. An existing work, Generalized Normality Learning (GNL), addresses domain shifts by enforcing feature consistency through training-time augmentation, but its reliance on prior knowledge of target distributions and access to training data at inference limits flexibility. To overcome these limitations, we propose a memory bank-based anomaly detection method that avoids retraining or access to training data during inference. We improve the robustness to distribution shifts via distribution alignment based test-time training. Our approach leverages a modified Sinkhorn distance to align distributions and handle outliers, offering a more resilient solution for industrial anomaly detection under realistic constraints. Extensive evaluations on out-of-distribution anomaly detection benchmarks demonstrate the effectiveness.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4311
Loading