Vote3D-AD: Unsupervised point cloud anomaly localization via varied defect synthesis and differentiable vote-clustering

Existing three-dimensional (3D) anomaly detection approaches typically rely on reconstruction, external memory banks, or fixed-radius clustering and often fail to generalize to noisy, irregularly sampled industrial scans or to capture the full diversity of real defects. We present Vote3D-AD, a single-pass framework that trains only on defect-free data and addresses these gaps with two principal contributions. First, we introduce Varied Defect Synthesis (VDS), a saliency-guided pseudo-anomaly generator that produces diverse, physically plausible defects (bulges, dents, holes, cracks, and surface roughness) together with sensor-level degradations to narrow the synthetic-to-real gap. Second, we develop a vote-and-cluster architecture in which local geometric representations predict learned, scale-aware votes that encode both spatial and boundary cues, and a differentiable soft-assignment clustering module aggregates these votes into coherent anomaly regions without relying on fixed-radius grouping or external memory structures. We evaluated our method on the synthetic Anomaly-ShapeNet benchmark and a new industrial dataset using three metrics: point-level Area Under the Receiver Operating Characteristic curve (AUROC), object-level Area Under the Precision-Recall curve (AUPR), and F1-Score. On average across both benchmarks, our method improves point-level AUROC by 6.7%, AUPR by 10.1% and F1 by 11.2%, and improves object-level AUROC by 5.3%, AUPR by 3.8% and F1 by 5.4% over the strongest baseline, while maintaining inference speeds above 9 frames per second (FPS).