MC3D-AD: A Unified Geometry-aware Reconstruction Model for Multi-category 3D Anomaly Detection

Jiayi Cheng, Can Gao, Jie Zhou, Jiajun Wen, Tao Dai, Jin-Bao Wang

TL;DR

This paper introduces MC3D-AD, a novel unified framework for 3D anomaly detection that can handle multiple object categories with a single model. By leveraging geometry-aware masked attention and a transformer-based architecture, MC3D-AD achieves state-of-the-art performance, significantly improving efficiency and generalization compared to existing single-category methods.

What Problem Does This Paper Address?

The manufacturing industry relies heavily on anomaly detection (AD) to ensure product quality. While 2D AD has seen significant progress, the increasing demand for high-precision industrial products has shifted focus to 3D AD using point cloud data. However, a major bottleneck in current 3D AD methods is their task-specific nature. Researchers must train a separate, highly customized model for each object category, which is costly, inefficient, and limits generalization. This paper tackles this challenge by proposing a unified model capable of performing multi-category 3D anomaly detection and localization without requiring category-specific training.

Methods

MC3D-AD is built upon a geometry-aware reconstruction framework that aims to learn robust representations of normal 3D point clouds. The core innovations lie in three components:

  • Adaptive Geometry-Aware Masked Attention (AGMA): This module is designed to overcome the "identity shortcut" problem, where models might simply copy input features without true understanding. AGMA explicitly computes neighborhood geometry variations (normal vectors and curvature) to guide a masked attention mechanism. By strategically masking key features based on these geometric variations, AGMA enhances the model's ability to learn intrinsic features and improves interpretability.
  • Local Geometry-Aware Encoder (LGE): Anomalies are often localized. The LGE leverages the AGMA module to encode local geometric information from point groups. It uses a transformer architecture with point cloud position embeddings and feature jittering to promote robust feature learning.
  • Global Query Decoder (GQD): To ensure accurate and complete anomaly localization, the GQD incorporates global information. It uses position embeddings as global queries and interacts with local features from the LGE through multiple AGMA layers. This global-local interaction facilitates feature fusion and improves decoding and reconstruction capabilities.

The overall framework reconstructs normal point cloud data. Anomalies are then detected by comparing the reconstruction error between the input and the reconstructed point clouds.

Datasets

The proposed MC3D-AD model was evaluated on two publicly available datasets:

  • Real3D-AD: This dataset features high-resolution point clouds from 12 object categories, with limited training samples per category and diverse anomaly shapes and sizes.
  • Anomaly-ShapeNet: This dataset is larger and more challenging, containing 1,600 samples across 40 categories with anomalies occupying 1-10% of the point cloud.

Key Insights and Ideas

  • Unified Multi-Category AD is Achievable: The paper demonstrates that a single model can effectively learn to detect anomalies across diverse object categories, challenging the prevailing paradigm of category-specific training.
  • Geometry is Key for Robustness: Explicitly capturing and utilizing local and global geometric variations is crucial for learning meaningful representations of normal point clouds, especially in the face of complex and disordered 3D data.
  • Masked Attention Guided by Geometry: The AGMA module's ability to intelligently mask features based on geometric variations is a novel approach to prevent shortcut learning and improve representation quality.
  • Transformer Architecture for 3D: The successful application of a transformer-based architecture, enhanced with geometric awareness and global query mechanisms, highlights its potential for complex 3D point cloud analysis tasks.

Key Results

MC3D-AD achieved state-of-the-art performance on both datasets:

  • Real3D-AD: Achieved an object-level AUROC of 0.782 and pixel-level AUROC of 0.768, outperforming the best single-category method by 3.1% and 1.0%, respectively.
  • Anomaly-ShapeNet: Demonstrated significant superiority with an object-level AUROC of 0.842 (9.3% improvement over the second-best) and pixel-level AUROC of 0.748 (8.0% improvement).

These results underscore the strong generalization capabilities of MC3D-AD across a wide range of object categories.

Limitations and Future Work

The authors acknowledge that while their method shows significant improvement, further research is needed. Specifically, they suggest exploring:

  • More Robust Frameworks: Developing even more robust and efficient frameworks for multi-category anomaly detection.
  • Exploiting Geometric Variation: Further investigating how to best utilize geometric variation information for anomaly detection.

The paper also implicitly highlights the challenge of handling extremely large and diverse datasets, which could be a direction for future work. The current approach relies on pre-trained models (PointMAE), and exploring self-supervised pre-training specifically for multi-category 3D AD could be another avenue.

Summary updated

Abstract

3D Anomaly Detection (AD) is a promising means of controlling the quality of manufactured products. However, existing methods typically require carefully training a task-specific model for each category independently, leading to high cost, low efficiency, and weak generalization. Therefore, this paper presents a novel unified model for Multi-Category 3D Anomaly Detection (MC3D-AD) that aims to utilize both local and global geometry-aware information to reconstruct normal representations of all categories. First, to learn robust and generalized features of different categories, we propose an adaptive geometry-aware masked attention module that extracts geometry variation information to guide mask attention. Then, we introduce a local geometry-aware encoder reinforced by the improved mask attention to encode group-level feature tokens. Finally, we design a global query decoder that utilizes point cloud position embeddings to improve the decoding process and reconstruction ability. This leads to local and global geometry-aware reconstructed feature tokens for the AD task. MC3D-AD is evaluated on two publicly available Real3D-AD and Anomaly-ShapeNet datasets, and exhibits significant superiority over current state-of-the-art single-category methods, achieving 3.1\% and 9.3\% improvement in object-level AUROC over Real3D-AD and Anomaly-ShapeNet, respectively. The code is available at https://github.com/iCAN-SZU/MC3D-AD.
BibTeX
@inproceedings{cheng2025mc,
  author = {Jiayi Cheng and Can Gao and Jie Zhou and Jiajun Wen and Tao Dai and Jin-Bao Wang},
  title = {{MC3D-AD: A Unified Geometry-aware Reconstruction Model for Multi-category 3D Anomaly Detection}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=XtikMpsd3a},
}
Figure