Reproduce “SplatPose & Detect: Pose-Agnostic 3D Anomaly Detection” on OpenResearch

About this paper

Mathis Kruse, Marco Rudolph, Dominik Woiwode, Bodo Rosenhahn

Researchers at Leibniz University Hannover developed SplatPose & Detect, a framework leveraging 3D Gaussian Splatting for pose-agnostic 3D anomaly detection that overcomes limitations of prior NeRF-based approaches. It achieves state-of-the-art detection and segmentation on the MAD dataset with an average AUROC of 93.9% and 99.5% respectively, while being significantly faster (55x faster training, 13x faster inference) and more robust to sparse data.

Key insights

• 3D Gaussian Splatting offers drastic computational efficiency gains (55x faster training, 13x faster inference) over NeRFs for 3D anomaly detection while maintaining or improving accuracy.
• Differentiably transforming the 3D Gaussian model itself to align with a query image's unknown pose is a robust and effective approach for fine pose estimation, outperforming iNeRF.
• The inherent robustness of 3D Gaussian Splatting to sparse training data makes it a more practical solution for real-world industrial applications where dense multi-view datasets are often unavailable.

Summary by alphaXiv.