Digital scanning devices are capable of acquiring high-resolution 3D models and have recently become affordable and commercially available. Modeling detailed 3D shapes by scanning real physical models is becoming commonplace. Current scanners are able to produce very large amounts of raw, dense point sets; consequently, there has been a recent increase in the need for techniques for processing point sets. One of the principal challenges faced today is the development of surface reconstruction techniques that deal with the inherent noise of the acquired dataset. When the underlying surface contains sharp features, the requirement of being resilient to noise is especially challenging, since noise and sharp features are ambiguous, and most existing techniques tend to smooth important features or even to amplify noisy samples. The project helps to train young researchers to work at the intersection of graphics, geometry, and statistics, while enabling them to pursue theoretically sound work that has deep practical impact. The data, software, and models developed in this project will be disseminated for other researchers to use in benchmarking and testing.
This research involves producing efficient and theoretically sound techniques for robust, feature-preserving surface reconstruction. It builds on recent work on the construction of a manifold surface from a set of points by using a moving least-squares (MLS) technique. The project explores the use of robust statistical techniques arising from outlier identification in MLS-based surface reconstruction. The approach is related to recent developments in feature-preserving smoothing, but it defines a surface rather than filtering the geometry. The techniques not only point to more reliable MLS projection but also extend the representation power of the underlying MLS surface definition to enable the representation of objects with sharp features.
