Belief propagation and self-adaptive voting dense stereo disparity estimation for textureless scene 3D Reconstruction

Abstract

High quality dense disparity map building was an important topic in diverse stereo vision research and applications. In practice, the textureless regions were always existed in both indoor and outdoor environment, and still not be well done for stereo vision system always with some drawbacks areas in dense disparity map. For more widely usability of various lenses, a toilless distortion correction was achieved by combing with a proposed autonomous calibration method and polar constraints. A color similarity probability based belief propagation algorithm (BP) is proposed to solve the depth discontinuous problem of occlusion and obtain an initial dense disparity map with kinds of drawbacks. In accurate disparity estimation step, Mean-Shift algorithm was utilized for initial similar disparity plane segment and a self-adaptive voting disparity plane fitting method was proposed for fitting optimization to adjust the drawbacks in the initial disparity estimation. Experimental results of real images in both indoor and outdoor environments validate the efficiency.