Depth estimation via light field camera with a hybrid stereo matching method

Abstract

Light field camera has the ability to capture ray's position and direction information through a single shot. The multiview images generated from the light field image provides unique advantage for depth estimation. In this paper, a method for 4D light field depth estimation was proposed by exploiting a hybrid stereo matching method with multidirectional multi-baseline stereo matching and K-Means segmentation. The multi-baseline stereo matching is adopted to effectively utilize the information among multiple low resolution sub-images. The multi-directional consistency checking is used to identify the fallible matching results due to occlusion or lack of texture. The K-Means segmentation method is used for locate the object boundary and hence ensure that the match window lies within a same object. A weighted median filtering method is adopted to further increase the quality of the disparity map. The experiment verifies the effectiveness of the method used. We evaluated the performance of our method on a publicly available dataset. Result shows that the proposed method can achieve a competitive result with a reasonable running time.