Dense Depth Map Reconstruction from Sparse Measurements Using a Multilayer Conditional Random Field Model

摘要: Acquiring accurate dense depth maps is crucial for 3D reconstruction. Current high quality sensors capable of generating are expensive and bulky, while compact low-cost can only reliably generate sparse measurements. We propose a novel multilayer conditional random field (MCRF) approach to reconstruct map target scene given the measurements corresponding photographic obtained from stereo photogrammetric systems. Estimating formulated as maximum posterior (MAP) inference problem where smoothness prior assumed. Our MCRF model uses measurement an additional observation layer describes relations between nodes with multivariate feature functions based on The method first qualitatively analyzed when performed data collected camera, then quantitative performance measured using Middlebury vision ground truth. Experimental results show our performs well reconstructing simple scenes has lower mean squared error compared other reconstruction methods.