A method for improving depth estimation in light field images

Abstract

The intensity of the light observed from every position and direction in a real scene can be modeled as a high-dimensional field, namely the plenoptic function. This field codes the radiance information as a function of space, orientation, wavelength, and time. In the scope of depth estimation, several strategies have been developed to obtain a representation of the spatial structure of a scene. However, existing methods do not take full advantage of the radiance information, such as edges, color, and texture. In this work, we propose a methodology for improving the estimation of depth maps in light field images by using segmentation and stereo matching algorithms. In this work, we apply classical image segmentation algorithms on the radiance image in order to obtain a detailed contour of the objects in the scene. Subsequently, a framework that unifies the results of image segmentation with depth estimation algorithms allows for improving the accuracy of the depth map. To validate the proposed methodology, two publicly available light field dataseis were used. The effectiveness of the proposed methodology is demonstrated through challenging real-world examples and including comparisons with the performance of state-of-the-art depth estimation algorithms.