Sparse Modeling of Shape from Structured Light

Abstract

Structured light depth reconstruction is among the most commonly used methods for 3D data acquisition. Yet, in most structured light methods, modeling of the acquired scene is crude, and is executed separately from the decoding phase. Here, we bridge this gap by viewing the reconstruction process via a probabilistic model combining illumination and shape. Specifically, an alternating minimization algorithm for structured light reconstruction is presented, incorporating a sparsity-based prior for the local surface model. Integrating this 3D surface prior into a probabilistic view of the reconstruction phase results in a robust estimation of the scene depth. We formulate and minimize reconstruction error and demonstrate performance of the algorithm on data from a structured light scanner. The results demonstrate the robustness of our algorithm to scanning artifacts under low SNR conditions and object motion.