PO3AD: Predicting Point Offsets toward Better 3D Point Cloud Anomaly Detection

TL;DR

PO3AD addresses 3D point cloud anomaly detection in the anomaly-free setting by shifting from reconstruction to per-point offset prediction on pseudo anomalies. The method introduces Norm-AS, a normal-vector guided augmentation that produces credible pseudo anomalies, enabling effective distillation of normal representations. A MinkUNet-based backbone with an offset predictor is trained using a joint offset loss that combines distance and direction terms to learn both normal and pseudo-abnormal offsets, producing anomaly scores directly from offsets. On Anomaly-ShapeNet and Real3D-AD, PO3AD achieves state-of-the-art results with average AUC-ROC gains of $9.0\%$ and $1.4\%$, respectively, demonstrating strong detection/localization performance and robustness to noise. The work offers a practical, data-efficient approach for 3D anomaly detection without relying on real anomalous samples, with potential for extension to multi-category models.

Abstract

Point cloud anomaly detection under the anomaly-free setting poses significant challenges as it requires accurately capturing the features of 3D normal data to identify deviations indicative of anomalies. Current efforts focus on devising reconstruction tasks, such as acquiring normal data representations by restoring normal samples from altered, pseudo-anomalous counterparts. Our findings reveal that distributing attention equally across normal and pseudo-anomalous data tends to dilute the model's focus on anomalous deviations. The challenge is further compounded by the inherently disordered and sparse nature of 3D point cloud data. In response to those predicaments, we introduce an innovative approach that emphasizes learning point offsets, targeting more informative pseudo-abnormal points, thus fostering more effective distillation of normal data representations. We also have crafted an augmentation technique that is steered by normal vectors, facilitating the creation of credible pseudo anomalies that enhance the efficiency of the training process. Our comprehensive experimental evaluation on the Anomaly-ShapeNet and Real3D-AD datasets evidences that our proposed method outperforms existing state-of-the-art approaches, achieving an average enhancement of 9.0% and 1.4% in the AUC-ROC detection metric across these datasets, respectively.

Figures (10)

• Figure 1: Comparison of reconstruction-based method and our method in terms of performance, and model attention. (a) Detection and localization performance of the reconstruction-based method on the ashtray0 category with various normal point loss weights; pseudo-abnormal points consistently weighted at 1.0 (implemented with our network due to the absence of official code). (b)By reducing normal weight in the reconstruction-based method, the model pays more attention to pseudo-abnormal points (marked with blue circles). Our method successfully focuses on pseudo-abnormal points. The model attention map is obtained by calculating the gradient of each point during backward propagation.
• Figure 2: Model structure comparison. (a) Restores normal samples from pseudo-abnormal variants; anomaly scores from input-output comparison. (b) Predicts point offsets of pseudo anomalies; anomaly scores from predicted offsets during testing.
• Figure 3: Illustration of our framework. Norm-AS generates pseudo anomalies from training normal samples. The backbone extracts features from pseudo anomalies, and the offset predictor estimates offsets for each point of input. The network trains under an offset loss constraint. During inference, the predicted offset distances serve as anomaly scores for test instances.
• Figure 4: Visualization of pseudo samples with and without normal vectors on the bottle0 category. Samples generated with normal vectors better mimic real anomalies.
• Figure 5: Detection and localization performance vs. patch no.