Open-Set Supervised 3D Anomaly Detection: An Industrial Dataset and a Generalisable Framework for Unknown Defects

Abstract

Although self-supervised 3D anomaly detection assumes that acquiring high-precision point clouds is computationally expensive, in real manufacturing scenarios it is often feasible to collect a limited number of anomalous samples. Therefore, we study open-set supervised 3D anomaly detection, where the model is trained with only normal samples and a small number of known anomalous samples, aiming to identify unknown anomalies at test time. We present Open-Industry, a high-quality industrial dataset containing 15 categories, each with five real anomaly types collected from production lines. We first adapt general open-set anomaly detection methods to accommodate 3D point cloud inputs better. Building upon this, we propose Open3D-AD, a point-cloud-oriented approach that leverages normal samples, simulated anomalies, and partially observed real anomalies to model the probability density distributions of normal and anomalous data. Then, we introduce a simple Correspondence Distributions Subsampling to reduce the overlap between normal and non-normal distributions, enabling stronger dual distributions modeling. Based on these contributions, we establish a comprehensive benchmark and evaluate the proposed method extensively on Open-Industry as well as established datasets including Real3D-AD and Anomaly-ShapeNet. Benchmark results and ablation studies demonstrate the effectiveness of Open3D-AD and further reveal the potential of open-set supervised 3D anomaly detection.

Figures (4)

• Figure 1: Comparison of 3D anomaly detection settings. Prior methods are unsupervised or rely on synthetic negatives without real negative-sample constraints, the model lacks a clear optimization direction and has no reliable notion of the negative feature distribution, leading to weak anomaly awareness. propose open-set supervised 3D anomaly detection, where the training set includes a few real anomalies from a subset of categories, enabling the model to learn a grounded notion of the negative feature distribution, while testing requires detecting unseen anomaly categories.
• Figure 2: Overview of the Open-Industry, comprising 15 categories, each containing five different anomalies. The red box represents scar, the green box represents scratch, the purple box represents concave, the brown box represents deformation, and the yellow box represents convex. The acquisition system and communication interaction are shown on the right.
• Figure 3: Workflow of our Open3D-AD. It consists of three steps: (i) anomaly synthesis and point-wise feature encoding to form normal/anomalous mixture representations, (ii) greedily subsample the normal support and use it to refine the anomalous support via deconfounding, and (iii) dual-distribution aggregated scoring for anomaly detection.
• Figure 4: Feature distribution visualizations in the training and inference stages. A and B show selected training features of Normal, Seen Abnormal, and Simulated Abnormal. C and D show selected test features of Normal, Seen Abnormal, and Unseen Abnormal. SS denotes the Silhouette Score. In A and B, $SS_1$, $SS_2$, and $SS_3$ correspond to Normal vs. Seen Abnormal, Normal vs. Simulated Abnormal, and Seen Abnormal vs. Simulated Abnormal, respectively. In C and D, they correspond to Normal vs. Seen Abnormal, Normal vs. Unseen Abnormal, and Seen Abnormal vs. Unseen Abnormal, respectively.