View Synthesis from Wide-Baseline Views Using Occlusion Aware Estimation of Large Disparities

Accurate disparity estimation is a key ingredient required when generating a high fidelity novel view from a set of input views. In this paper, a high quality disparity estimation method is proposed for view synthesis from multiple input images with large disparities and occlusions. The method optimally selects one out of three image pairs to estimate the disparity map for different regions of the novel view. The novel view is then formed using this disparity map. We introduce two novel elements: a) an enhanced visibility map that is able to segment the scene accurately near object boundaries and b) a backward unilateral and bilateral disparity estimation procedure using the Gabor transform on an expandable search window to tackle large disparities. The quality of the interpolated virtual views produced by the proposed method is assessed and compared against two of the prominent previously-reported methods. The proposed method offers a significant improvement both in terms of visual quality of the interpolated views as well as the peak signal-to-noise ratio (PSNR) and structured similarity image index (SSIM) metrics.