Depth from Defocus via Active Multispectral Quasi-random Point Projections using Deep Learning

A novel approach for inferring depth measurements via multispectral active depth from defocus and deep learning has been designed, implemented, and successfully tested. The scene is actively illuminated with a multispectral quasi-random point pattern, and a conventional RGB camera is used to acquire images of the projected pattern. The projection points in the captured image of the projected pattern are analyzed using an ensemble of deep neural networks to estimate the depth at each projection point. A final depth map is then reconstructed algorithmically based on the point depth estimates. Experiments using different test scenes with different structural characteristics show that the proposed approach can produced improved depth maps compared to prior deep learning approaches using monospectral projection patterns.