Bridging 3D Anomaly Localization and Repair via High-Quality Continuous Geometric Representation

TL;DR

This work introduces Pose-Aware Signed Distance Function (PASDF), a continuous, pose-invariant 3D representation that jointly performs anomaly detection and in-situ repair. By disentangling pose from shape via a Pose-wise Alignment Module and learning an SDF network, PASDF achieves state-of-the-art object- and pixel-level anomaly localization and can generate high-fidelity repair templates using marching cubes. Extensive experiments on Real3D-AD and Anomaly-ShapeNet demonstrate robust performance gains and the critical role of pose normalization and positional encoding. The approach offers a practical path toward reliable 3D anomaly detection with actionable repair capabilities in real-world applications.

Abstract

3D point cloud anomaly detection is essential for robust vision systems but is challenged by pose variations and complex geometric anomalies. Existing patch-based methods often suffer from geometric fidelity issues due to discrete voxelization or projection-based representations, limiting fine-grained anomaly localization. We introduce Pose-Aware Signed Distance Field (PASDF), a novel framework that integrates 3D anomaly detection and repair by learning a continuous, pose-invariant shape representation. PASDF leverages a Pose Alignment Module for canonicalization and a SDF Network to dynamically incorporate pose, enabling implicit learning of high-fidelity anomaly repair templates from the continuous SDF. This facilitates precise pixel-level anomaly localization through an Anomaly-Aware Scoring Module. Crucially, the continuous 3D representation in PASDF extends beyond detection, facilitating in-situ anomaly repair. Experiments on Real3D-AD and Anomaly-ShapeNet demonstrate state-of-the-art performance, achieving high object-level AUROC scores of 80.2% and 90.0%, respectively. These results highlight the effectiveness of continuous geometric representations in advancing 3D anomaly detection and facilitating practical anomaly region repair. The code is available at https://github.com/ZZZBBBZZZ/PASDF to support further research.

Figures (5)

• Figure 1: (a) 3D representation comparison for anomaly detection: voxel-based, point-based, projection-based, and our approach with signed distance function (SDF). (b) PASDF's continuous geometric representation enables both high-precision anomaly localization and effective anomaly surface repair.
• Figure 2: Overview of Pose-Aware Signed Distance Function (PASDF) for 3D Anomaly Detection. A normal point cloud is selected as the Canonical Pose, and all training and test point clouds are aligned to it via the Pose-wise Alignment Module. Query points are categorized as surface ( ), outside ( ), or inside ( ). Their coordinates, enriched with positional encoding, form the feature $z$, which serves as input to the SDF network $\phi$. During training, a clamped L1 loss is applied, while at inference, deviations from learned SDF values are used for anomaly scoring.
• Figure 3: Qualitative anomaly localization results on Anomaly-ShapeNet and Real3D-AD (rows 1-2 and 3-4 respectively).
• Figure 4: Qualitative evaluation of PASDF for reconstruction. The first three columns are from the Real3D AD dataset, and the last three columns are from Anomaly-ShapeNet. w/o $PE$ denotes PASDF without positional encoding
• Figure A: Qualitative anomaly localization results on Anomaly-ShapeNet.