Structured Light Based Reconstruction under Local Spatial Coherence Assumption

Abstract

3D scanning techniques based on structured light usually achieve robustness against outliers by performing multiple projections to simplify correspondence. However, for cases such as dynamic scenes, the number of frames captured from a certain view must be kept as low as possible, which makes it difficult to reconstruct complex scenes with high frequency shapes and inappropriate reflection properties. To tackle this problem, we present a novel set of color stripe patterns and a robust correspondence algorithm that assume local spatial coherence in the captured data. This assumption allows us to design our stripe sequences with globally unique neighborhood properties to effectively avoid wrong correspondences. The concept of local spatial coherence is further exploited to make the ensuing surface reconstruction practically insensitive to noise, outliers, and anisotropic sampling density. Thus, the recovery of a topologically consistent manifold surface can be drastically simplified. We have successfully generated high quality meshes of various colored objects using a minimalistic projector-camera system. In particular, the full sampling capabilities of our devices can be exhausted by taking only three shots.