Probing 3D anomalies via multi-view registration and dual-residual analysis

Due to the sparse, unordered, and low information-density characteristics of point cloud data, 3D anomaly detection is a challenging task, especially when there is only a small amount of normal training data. Recently, registration-based point cloud anomaly detection has gained increasing popularity. To address the inherent local optima and random initial posture issues in registration-based methods, we revisit registration from a multi-view perspective. Then, a downsampled coarse-to-fine registration pipeline is utilized to align the test samples with the enhanced normal references. Geometric and feature residuals are combined to capture both point-wise spatial deviations and differences in the high-dimensional feature space distribution. Our framework is evaluated on public Anomaly-ShapeNet and Real3D-AD datasets, along with other state-of-the-art works at the object and point levels. A detailed ablation study is conducted to analyze the effects of the multi-view generation strategy, dual-branch residuals, and input features. Extensive experiments validate the effectiveness of our proposed method, achieving improvements of 4.1% and 1.8% in object-level AUROC, respectively. The code will be released to support further research.