Empowered by advanced plenoptic sensing systems, light-field imaging becomes one of the most extensively used methods for capturing 3D views of a scene. In contrast to the traditional input to a 3D graphics system, namely, scenes consisting of pre-defined geometric primitives with different materials and sets of lights, the input to a light field is only a set of 2D images which are informative and cost effective. Unfortunately, due to the limited sensor resolution, existing systems must balance the spatial and angular resolution, i.e., one can obtain dense sampling images in the spatial dimension but only sparse sampling images in the angular (viewing angle) dimension or vice versa.