Registration is a Powerful Rotation-Invariance Learner for 3D Anomaly Detection

TL;DR

This work tackles 3D anomaly detection in point clouds under rotation and partial overlap by integrating point-cloud registration into feature learning. The proposed Reg2Inv framework jointly trains a registration-guided, rotation-invariant feature extractor and a memory-bank-based anomaly detector, using multi-scale patches and optimal transport objectives to align and compare normal prototypes with test samples. It demonstrates superior performance on Real3D-AD and Anomaly-ShapeNet, delivering robust localization and improved discriminability by emphasizing local geometric detail and transformation consistency. The method holds practical value for industrial QA by providing more reliable defect detection across orientations, with ablations confirming the importance of the registration loss, feature normalization, and memory-bank composition. Limitations include per-class training costs and potential registration inaccuracies, guiding future work to improve scalability and robustness.

Abstract

3D anomaly detection in point-cloud data is critical for industrial quality control, aiming to identify structural defects with high reliability. However, current memory bank-based methods often suffer from inconsistent feature transformations and limited discriminative capacity, particularly in capturing local geometric details and achieving rotation invariance. These limitations become more pronounced when registration fails, leading to unreliable detection results. We argue that point-cloud registration plays an essential role not only in aligning geometric structures but also in guiding feature extraction toward rotation-invariant and locally discriminative representations. To this end, we propose a registration-induced, rotation-invariant feature extraction framework that integrates the objectives of point-cloud registration and memory-based anomaly detection. Our key insight is that both tasks rely on modeling local geometric structures and leveraging feature similarity across samples. By embedding feature extraction into the registration learning process, our framework jointly optimizes alignment and representation learning. This integration enables the network to acquire features that are both robust to rotations and highly effective for anomaly detection. Extensive experiments on the Anomaly-ShapeNet and Real3D-AD datasets demonstrate that our method consistently outperforms existing approaches in effectiveness and generalizability.

Figures (7)

• Figure 1: (a) Existing methods still exhibit misalignments after registration. (b) Visualization of PointMAE features in the memory bank, with different colors for different objects. (c) Our method achieves accurate point cloud alignment. (d) Visualization of our locally discriminative features in the memory bank.
• Figure 2: The overview of Reg2Inv. In the feature learning stage, the model learns features via a registration task that enforces geometric alignment and multi-scale consistency between paired point clouds. At inference time, the model computes a registration matrix for alignment with the prototype and extracts rotation-invariant features for anomaly detection. The feature extractor is designed to obtain a set of features for anomaly detection and two sets of features for registration.
• Figure 3: Visualization comparison of localization results on Real3D-AD. (a) Ground truth. (b) Ours. (c)&(e) PatchCore using FPFH and PointMAE, respectively. (d) ISMP. (f) Reg3D-AD.
• Figure 4: Ablation study on neighborhood sizes.
• Figure 5: More qualitative results of localization on the Real3D-AD dataset.