R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection

TL;DR

This paper proposes R3D-AD, reconstructing anomalous point clouds by diffusion model for precise 3D anomaly detection, which outperforms previous state-of-the-art methods and presents a novel 3D anomaly simulation strategy named Patch-Gen to generate realistic and diverse defect shapes.

Abstract

3D anomaly detection plays a crucial role in monitoring parts for localized inherent defects in precision manufacturing. Embedding-based and reconstruction-based approaches are among the most popular and successful methods. However, there are two major challenges to the practical application of the current approaches: 1) the embedded models suffer the prohibitive computational and storage due to the memory bank structure; 2) the reconstructive models based on the MAE mechanism fail to detect anomalies in the unmasked regions. In this paper, we propose R3D-AD, reconstructing anomalous point clouds by diffusion model for precise 3D anomaly detection. Our approach capitalizes on the data distribution conversion of the diffusion process to entirely obscure the input's anomalous geometry. It step-wisely learns a strict point-level displacement behavior, which methodically corrects the aberrant points. To increase the generalization of the model, we further present a novel 3D anomaly simulation strategy named Patch-Gen to generate realistic and diverse defect shapes, which narrows the domain gap between training and testing. Our R3D-AD ensures a uniform spatial transformation, which allows straightforwardly generating anomaly results by distance comparison. Extensive experiments show that our R3D-AD outperforms previous state-of-the-art methods, achieving 73.4% Image-level AUROC on the Real3D-AD dataset and 74.9% Image-level AUROC on the Anomaly-ShapeNet dataset with an exceptional efficiency.

Figures (7)

• Figure 1: Comparison of architectures. (a) Embedded model encodes the input $\mathcal{X}$ into features and stores them in the memory bank during training. The anomaly map $\mathcal{M}$ is obtained by comparing the test features with all the features in the memory bank. (b) Reconstructive model is trained by minimizing the loss between its input $\mathcal{X}$ and the output $\widehat{\mathcal{X}}$. The anomaly map $\mathcal{M}$ is obtained by comparing the test phase input with its corresponding reconstruction target.
• Figure 2: Overall architecture of R3D-AD for shape reconstruction and anomaly detection of point cloud objects. Reconstruction training phase: The simulated anomalous $\mathcal{P}_\text{a}^{(0)}$ is generated resort to Patch-Gen from the input point cloud. It further fully masked as $\mathcal{P}_\text{a}^{(T)}$ while also encoded to latent shape embedding. The SWD decoder then explicitly reconstructs the anomaly-free object $\mathcal{P}_\text{r}^{(0)}$ with consistent spatial transform by conditionally generating point-level displacements $\mathrm{\Delta} ^{(t)}$ at each step of the inverse process. Detection testing phase: The test point cloud $\mathcal{P}_\text{a}^{(0)}$ is reconstructed to $\mathcal{P}_\text{r}^{(0)}$ with normal shape, and compared at a distance level to detect the anomalous region.
• Figure 3: Illustration of Patch-Gen, the 3D anomaly simulation strategy. The input normal point cloud is first randomly rotated. On the surface of the normalized cube, we randomly select viewpoints to find the nearest patch of points. The selected points are then transformed into irregular defects according to the specific deformation solution.
• Figure 4: Memory and time cost during inference on Real3D-AD dataset. (a) Memory usage comparison between different models. (b) 3D anomaly detection performance vs. frames per second on an NVIDIA RTX 3090 GPU. Our R3D-AD outperforms all previous methods on both accuracy and efficiency by a significant margin.
• Figure 5: Qualitative analysis on Real3D-AD dataset and Anomaly-ShapeNet dataset. The anomaly map is obtained directly by calculating the differences between the input and reconstructed point clouds, where deeper colors represent more confidence.