Abstract
A key step in any scanning-based asset creation workflow is to convertunordered point clouds to a surface. Classical methods (e.g., Poissonreconstruction) start to degrade in the presence of noisy and partial scans.Hence, deep learning based methods have recently been proposed to producecomplete surfaces, even from partial scans. However, such data-driven methodsstruggle to generalize to new shapes with large geometric and topologicalvariations. We present Points2Surf, a novel patch-based learning framework thatproduces accurate surfaces directly from raw scans without normals. Learning aprior over a combination of detailed local patches and coarse globalinformation improves generalization performance and reconstruction accuracy.Our extensive comparison on both synthetic and real data demonstrates a clearadvantage of our method over state-of-the-art alternatives on previously unseenclasses (on average, Points2Surf brings down reconstruction error by 30\% overSPR and by 270\%+ over deep learning based SotA methods) at the cost of longercomputation times and a slight increase in small-scale topological noise insome cases. Our source code, pre-trained model, and dataset are available on:https://github.com/ErlerPhilipp/points2surf