A Semantically Disentangled Unified Model for Multi-category 3D Anomaly Detection

TL;DR

This paper introduces a novel semantically disentangled unified model for 3D anomaly detection in point clouds, improving accuracy and reliability by addressing inter-category feature entanglement.

Contribution

It proposes a new framework with semantic disentanglement components, enhancing unified 3D anomaly detection across multiple categories.

Findings

Achieves state-of-the-art AUROC improvements on Real3D-AD and Anomaly-ShapeNet datasets.

Effectively disentangles category semantics, reducing inter-category entanglement.

Enhances the reliability of unified 3D anomaly detection models.

Abstract

3D anomaly detection targets the detection and localization of defects in 3D point clouds trained solely on normal data. While a unified model improves scalability by learning across multiple categories, it often suffers from Inter-Category Entanglement (ICE)-where latent features from different categories overlap, causing the model to adopt incorrect semantic priors during reconstruction and ultimately yielding unreliable anomaly scores. To address this issue, we propose the Semantically Disentangled Unified Model for 3D Anomaly Detection, which reconstructs features conditioned on disentangled semantic representations. Our framework consists of three key components: (i) Coarse-to-Fine Global Tokenization for forming instance-level semantic identity, (ii) Category-Conditioned Contrastive Learning for disentangling category semantics, and (iii) a Geometry-Guided Decoder for semantically consistent reconstruction. Extensive experiments on Real3D-AD and Anomaly-ShapeNet demonstrate that our method achieves state-of-the-art for both unified and category-specific models, improving object-level AUROC by 2.8% and 9.1%, respectively, while enhancing the reliability of unified 3D anomaly detection.