A Uniform Bayesian Framework for Integration

Vision researchers have advocated the integration of vision modules. However, generic system integration issues for recovering 3D information have not been adequately addressed in the literature. We propose a uniied Bayesian integration framework for interactions among the vision modules to obtain a complete 3D reconstruction from a pair of intensity (stereo) images. We integrate perceptual grouping, stereo, shape from shading, and shape from texture modules under the proposed framework and demonstrate that the integrated system recovers the depth and surface orientation information more reliably than the individual modules for diierent synthetic and real images. Inferring intrinsic properties (depth, orientation, albedo, and reeectance) of the physical surfaces from their intensity images is one of the central problems in computer vision. During last two decades, a number of algorithms have been developed for extracting the 3D information from the intensity images using the individual visual cues (e.g., stereo, texture, shading) and have been commonly referred to as vision modules. These individual vision modules independently can not obtain accurate 3D reconstruction of the scene due to several 1