Point Cloud Diffusion with Global and Local Reconstruction for Instance-Level 3D Anomaly Detection

In high-precision industrial manufacturing, detecting subtle 3D anomalies (such as fine scratches) in point clouds is challenging due to the extreme similarity between defective and normal regions. Existing reconstruction-based methods often struggle with two issues: they fail to reconstruct small-scale, weak defects and frequently introduce positional bias in background regions, leading to false positives. The authors investigate how to improve anomaly detection accuracy by better modeling these subtle defects and ensuring structural consistency during reconstruction.

The authors propose PCDiff, a unified diffusion-based framework consisting of two primary phases: generation and detection.

1. Generation Phase: To overcome the lack of anomaly data, the authors introduce a texture augmentation module that synthesizes diverse, realistic defects by perturbing mesh vertices using a guidance bank of geometric primitives. They employ an instance-level multi-modal attention mechanism that conditions the diffusion process on texture gradients, image patches, text, and spatial masks.

2. Detection Phase: The authors implement a joint local-global reconstruction algorithm. A local branch uses 2D-derived anomaly masks (obtained via multi-view rendering and back-projection) to focus on restoring defective regions. Simultaneously, a global branch performs unbiased reconstruction of the entire point cloud. By combining these, the model preserves the background structure while effectively highlighting foreground anomalies.

This work addresses a critical bottleneck in automated visual inspection: the difficulty of detecting fine-grained defects that are nearly indistinguishable from normal geometric noise. By integrating 2D visual priors into 3D diffusion and employing a dual-branch reconstruction strategy, PCDiff achieves higher fidelity in anomaly generation and more precise detection than current state-of-the-art methods. This provides a more robust solution for quality control in high-density electronics and other precision manufacturing sectors.