A Method for Handling Multi-Occlusion in Depth Estimation of Light Field
Xihao Pan, Tao Zhang, Hao Wang · 2019
With the blossom of light field cameras, a number of studies followed. Depth estimation becomes one of the most important problems in light field and numerous algorithms have been proposed to achieve accurate depth map from a light field image. However, few algorithms take occlusions into consideration, especially for multi-occlusion. This makes incorrect depth estimation in object boundaries. In this paper, based on theories of the photo-consistency for Lambertian surfaces and the correspondence between spatial space and angular space, we proposed a method to handle occlusions in depth estimation of light field. The occlusion edges are extracted firstly by the Lucas-Kanade optical flow algorithm. Secondly K-means clustering is applied for each pixel in angular patch adaptively. Thirdly the initial depth map is obtained by aggregating the modified matching costs respectively in the angular patch. Finally the depth map is regularized using Markov Random Field and weighted median filter. Experimental results show that the proposed method achieves more accurate depth map on synthetic datasets, especially in multi-occlusion boundaries.