Disparity Estimation for Focused Light Field Camera Using Cost Aggregation in Micro-Images

Unlike conventional light field camera that records spatial and angular information explicitly, the focused light field camera implicitly collects angular samplings in micro-images behind the micro-lens array. Without directly decoded sub-apertures, it is difficult to estimate disparity for focused light field camera. On the other hand, disparity estimation is a critical step for sub-aperture rendering from raw image. It is hence a typical "chicken-and-egg" problem. In this paper we propose a two-stage method for disparity estimation from the raw image. Compared with previous approaches which treat all pixels in a micro-image as a same disparity label, a segmentation-tree based cost aggregation is introduced to provide a more robust disparity estimation for each pixel, which optimizes the disparity of low-texture areas and yields sharper occlusion boundaries. After sub-apertures are rendered from the raw image using initial estimation, the optimal one is globally regularized using the reference sub-aperture image. Experimental results on real scene datasets have demonstrated advantages of our method over previous work, especially in low-texture areas and occlusion boundaries.