Block-wise MAP disparity estimation for intermediate view reconstruction

Abstract

A dense disparity map is required in the application of intermediate view reconstruction from stereoscopic images. A popular approach to obtaining a dense disparity map is maximum a-posteriori (MAP) disparity estimation. The MAP approach requires statistical models for modeling both a likelihood term and an a-priori term. Normally, a Gaussian model is used. In this contribution, block-wise MAP disparity estimation using different statistical models are compared in terms of Peak Signal-to-Noise Ratio (PSNR) of disparity-compensation errors and number of corresponding matches. It was found that, among the Cauchy, Laplacian, and Gaussian models, the Laplacian model is the best for the likelihood term while the Cauchy model is the best for the a-priori term. Experimental results show that reconstruction algorithm with the MAP disparity estimation using the determined models can improve image quality of the intermediate views reconstructed from stereoscopic image pairs.