Gradient and Color Intensity based Dense Disparity Estimation Using Adaptive Weight Aggregation

Abstract

Dense disparity estimation is an important step in reconstruction of 3D scene which provides disparity at each pixel of an image. Disparity estimation techniques employ various attributes of an image pixel such as gray level, color intensity and gradient value for accurate depth computation. Both the intensity as well as the gradient based cost measures have their shortcomings: the intensity based methods are noise prone whereas the gradient based cost measures fail in smooth regions. In this paper, a disparity estimation scheme has been proposed which applies the gradient cost measure on grayscale and sum of absolute difference cost measure on color level which is aggregated using guided filtering. The disparity map refinement is carried out using weighted median filtering and a filtering mask. This scheme is applied to the standard Middlebury data set and compared online with other existing methodologies. Simulation results prove the efficacy of the proposed technique and is found to perform better than most of the existing techniques.