The automatic detection of surface-level irregularities—defects or anomalies—in 3D data is of significant interest for various real-world purposes, such as industrial quality inspection, infrastructure monitoring, robotics, and autonomous systems. However, collecting annotated defect examples at a large scale is costly, and existing 3D anomaly detection methods either require templates or heavy memory, multiple inference passes, and brittle heuristic clustering. These shortcomings limit real-life deployment.
In a novel research work, a team of researchers from Japan and Vietnam, led by Associate Professor Phan Xuan Tan, affiliated with the Innovative Global Program, College of Engineering, Shibaura Institute of Technology, Japan, and including Dr. Dinh-Cuong Hoang from FPT University, Vietnam, has proposed Vote3D-AD as a new single-pass framework for 3D anomaly localization. Their novel findings were made available online on January 20, 2026, and have been published in Volume 137 of the Alexandria Engineering Journal on February 1, 2026.
Vote3D-AD combines a realistic pseudo-anomaly generator (Varied Defect Synthesis, or VDS), with a learned vote-and-differentiable clustering architecture to localize defects in point clouds, while training only on defect-free data. It demonstrates stronger and more reliable localization than prior works. Across both synthetic and real industrial benchmarks, Vote3D-AD improves point-level AUROC by ~6.7%, and point-AUPR by ~10.1%, and point-F1 by ~11.2% over the strongest baselines. Notably, object-level metrics also rise substantially.
The proposed framework exhibits practical speed for real inspection pipelines. The full pipeline runs at approximately 9.05 FPS on an RTX-3090 and supports higher-throughput variants, demonstrating a balance of high accuracy and practical inference speed for production inspection. Vote3D-AD turns sparse, noisy point-level signals into coherent region proposals without hand-tuned clustering, reducing the engineering burden and improving localization precision, which is critical for making automated 3D inspection actionable on factory floors.
Its real-world applications include automated industrial quality inspection of sheet metal, machined parts, and plastic housings, where it detects dents, bulges, holes, missing components, and surface roughness from depth scans, especially in situations where RGB or 2D images miss geometric faults. Vote3D-AD is also expected to be useful for infrastructure and asset monitoring. It can localize cracks, holes, and surface degradation in scanned components, such as pipes, panels, and connectors, where early geometric anomalies predict failure. Moreover, it may enable robots to verify assembly quality or to detect damage on manipulated objects using onboard depth sensors, even when only defect-free examples are available during training. Furthermore, the framework can combine multi-view or streamed point clouds to detect thermal or structural anomalies—with potential future multi-modal fusion—improving predictive maintenance decisions.
According to Dr. Tan, "Tighter and more coherent anomaly masks present in Vote3D-AD—thanks to learned clustering and boundary refinement—reduce false alarms and ambiguous detections, lowering unneeded rework and downtime in manufacturing lines. In addition, since Vote3D-AD trains on normal examples only and uses learned vote clustering instead of template matching or big feature stores, it is easier to deploy across different product families and manufacturers. This reduces engineering and data collection costs."
"Furthermore, our technology utilizes VDS that generates diverse, physically plausible pseudo-defects such as bulges, dents, holes, cracks, and surface roughness, and simulates sensor artifacts such as noise and dropout, which substantially narrows the gap between simulated training signals and real defects, improving generalization," highlights Dr. Tan.
Lastly, by producing coherent region proposals rather than scattered point outliers, the proposed system supports downstream actions such as automatic rejection, targeted repair, and prioritization for human review, making inspection pipelines more efficient and safer.
