Noise is More Than Just Interference: Information Infusion Networks for Anomaly Detection
Keywords: Self-supervised learning, Anomaly detection
Abstract: 3D anomaly detection is a crucial task in computer vision, aiming to identify anomalous points or regions from point cloud data. However, existing methods may encounter challenges when handling point clouds with high intra-class variance, especially for methods that rely on registration techniques. In this study, we propose a novel 3D anomaly detection method, termed Information Gain Block-based Anomaly Detection (IGB-AD), to address the challenges of insufficient anomaly detection information and high intra-class variance. To extract ordered features from 3D point clouds, the technique of Rotation-Invariant Farthest Point Sampling (RIFPS) is first introduced. Then, an Information Perfusion (IP) module composed of stacked Information Gain Blocks (IGB) is proposed to utilize prior noise to provide more distinguishing information for the features, where IGB is designed to utilize noise in a reverse-thinking manner to enhance anomaly detection. Finally, a Packet Downsampling (PD) technique is developed to preserve key information between multiple clusters to solve the complex downsampling situation. The main purpose of the framework is to utilize the effective information within prior noise to provide more detection criteria for anomaly detection. In addition, an Intra-Class Diversity (ICD) 3D dataset is constructed, which contains multiple categories with high class-variance. Experimental results show that the proposed IGB-AD method achieves the State-Of-The-Arts (SOTA) performance on the Anomaly ShapeNet dataset, with an P-AUROC of 81.5% and I-AUROC of 80.9%, and also gains the best performance on the ICD dataset, with an P-AUROC of 57.4% and I-AUROC of 60.2%. Our dataset will be released after acceptance.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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: 739
Loading