Composite pattern separation with CNN for multi-view structured light 3D reconstruction

Structured light(SL) is a fundamental and important method in 3D reconstruction. Multi-view SL can provide broader and more complete scene geometry than single SL, but the overlap of projected patterns is a fatal problem, and thus it is very much desired to separate signals from the composite pattern. In this paper, we propose a novel learning-based method to decompose patterns for phase-coding SL. The network has a backbone to extract common features and two branches to get the unique features and reconstruct the two components. In addition, by analyzing the process of the pattern overlapping, we propose a novel loss function and ground truth label data by guaranteeing correct phase. Moreover, we also propose a synthetic dataset for training. Experiment results demonstrate that the proposed method successfully separate the composite patterns on our data with noise and light attenuation, and outperforms traditional method such as band pass filter and principal component analysis.