Surfel Based Geometry Resonstruction

We propose a method for retrieving a piecewise smooth surface from noisy data. In data acquired by a scanning process sampled points are almost never on the discontinuities making reconstruction of surfaces with sharp features difficult. Our method is based on a Markov Random Field (MRF) formulation of a surface prior, with the surface represented as a collection of small planar patches, the surfels, associated with each data point.

The main advantage of using surfels is that we avoid treating data points as vertices. MRF formulation of the surface prior allows us to separately model the likelihood (related to the mesh formation process) and the local surface properties. We chose to model the smoothness by considering two terms: the parallelism between neighboring surfels, and their overlap.

We have demonstrated the feasibility of this approach on both synthetical and scanned data. In both cases sharp features were precisely located and planar regions smoothed.