Wavelet-Based Stereo Vision

Abstract

Multiresolution frameworks have been embraced by the stereo imaging community because of their human-like approach in solving the correspondence problem and reconstructing density maps from binocular images. We describe a method to recover depth information of stereo images based on a multi-channel wavelet transform, where trends in the coefficients provide overall context throughout the framework, while transients are used to give refined local details into the image. A locally adapted lifting scheme is used to maximize the subband decorrelation energy by the transients. The coefficients in each channel computed from the lifting framework are combined to measure the local correlation of matching windows in the stereogram. The combined correlation yields higher cumulative confidence in the disparity measure than using a single primitive, such as LOG, which has been applied to the traditional area-based stereo techniques.