Wavelet-based Surface Reconstruction from Multi-Scale Sample Points

Multi-view stereo reconstruction techniques yield inherently multi-scale point data typically fed into surface reconstruction algorithms. Following the intuition of scale space we assume that sample points originate from smoothed versions of the original surface. The smoothing can be characterized by a smoothing kernel that suppresses fine-scale structures. In this paper, we propose a surface reconstruction framework that correctly handles this multi-scale input data. We represent the surface using a multi-resolution analysis allowing us to reconstruct scales separately and to merge the sample points in frequency space. With an underlying wavelet basis we are able to locally model surface detail according to the surface properties or sample distribution. We first demonstrate the effectiveness of our method on a synthetic data set with known smoothing. For real-world data obtained by multiview stereo we estimate the smoothing kernel and present reconstruction results with enhanced detail.