Robust disparity estimation on sparse sampled light field images

Abstract

The paper presents a robust approach to compute disparities on sparse sampled light field images based on Epipolar-Plane Image (EPI) analysis. The Relative Gradient is leveraged as a kernel density function to cope with radiometric changes in non-Lambertian scenes. To account for the sparse light field, a window-based filtering is introduced to handle the noisy and homogenous regions, decomposing the scene images into edge and non-edge regions. Separate score-volume filtering over these regions avoids boundary fattening effects common to stereo matching. Finally, a consistency measure detects unreliable pixels with false disparities, to which a disparity refinement is applied. Evaluation analysis is performed on the Disney light field dataset and the proposed method shows superior results over state-of-the-art.