| In this dissertation, I report the progress towards building a robust and efficient 3D reconstruction system based on stereo vision. Stereo vision is known to be quite fragile in practice due to specular highlights, lack of texture, lighting variations, image blurring, etc. In this dissertation, I focus on exploiting the relationships between illuminants, surface reflection and shape to increase the robustness of stereo vision. I first present a new image transform for matching low-textured regions and then a robust solution for illumination chromaticity estimation based on a new correspondence matching invariant called Illumination Chromaticity Constancy. I next propose a new framework based on bilateral filtering and loopy belief propagation for simultaneous estimation of surface reflectance and shape with the assumption that the illumination chromaticity can be correctly estimated. Two new bilateral filtering algorithms with computational complexity invariant to filter kernel size and a new belief propagation with computational complexity invariant to the disparity search range are then presented to reduce the speed and memory cost. |
