GPU-based depth estimation for light field images

With the development in light field imaging, some depth map estimation algorithms have been proposed for plenoptic cameras. They fuse the information from multiple depth cues to improve the estimation accuracy, while high computational complexity is also introduced, which is not applicable to real-time applications. Thus, in this paper, a GPU-based depth estimation is proposed for plenoptic images to obtain depth information accurately in a very short time. An angular patch-based thread allocation and depth-adaptive data-loading method are proposed to adapt the parallel characteristics of GPU to depth estimation and to reduce the data transfer bandwidth efficiently. Experimental results show that the proposed GPU optimization provides an average of more than 2000 times acceleration relative to the single thread implementation of depth estimation on CPU, which is very beneficial to real-time applications.