EG3AD: An Efficient Geometry-Aware Encoding Framework for Reconstruction-Based Multi-Class Point Cloud Anomaly Detection

Yuansong Wang, Yaohua Zha, Liao Wenwen, Chunlin Fan, Tao Dai, Shu-Tao Xia

Keywords: Point Cloud; Anomaly Detection; Reconstrction Model

TL;DR: We propose an efficient geometry-aware encoding framework for reconstruction-based multi-class point cloud anomaly detection.

Abstract:

Multi-class point cloud anomaly detection is a critical task that aims to identify anomalous patterns across various categories using a single, unified model.
Current reconstruction based methods predominantly rely on transformer encoders to extract high-level semantic features, aiming to filter out subtle defect features, and then use decoders to reconstruct them into normal patterns. However, this suffers from two limitations:
(1) employing encoders based on global attention mechanisms, particularly on uniformly tokenized inputs, hinders the rapid extraction of fine-grained local features;
(2) high computational cost arising from stacking multiple encoding layers during semantic feature extraction.
Thus, we propose EG3AD, an Efficient Geometry-aware encoding framework for reconstruction-based multi-class 3D point cloud Anomaly Detection.
To investigate how to obtain effective geometric representations under token and parameter constraints, we begin by introducing the Curvature-Aware Sampling module, which mitigates the distortions caused by uniform sampling in regions of high curvature.
Then, leveraging geometry prior bias of point cloud data, we design the Point Cluster Graph Convolution, which enables compact and effect local geometric aggregation through only limited lightweight layers.
Finally, to obtain anomaly-invariant semantic features without relying on deep encoding layers, we introduce the Feature Purification Module inspired by optimal transport theory. This module compresses features into a set of cluster centroids that preserve fundamental geometric patterns, thereby yielding representations robust to subtle anomalies.
Extensive experiments show that simply replacing the vanilla point transformer encoder with our proposed EG3AD yields state-of-the-art results on all PCAD benchmarks.
Our code will be made publicly available upon acceptance.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 3215