An optical analysis of light field rendering

Light field rendering has been proposed to render novel images from a collection of captured images without explicit geometry information. In this report, we present an optical analysis of light field rendering. In particular, a light field rendering system can be considered as a virtual imaging system with a discrete synthetic aperture (e.g., [10]. We define the key elements of the virtual imaging system in-cluding the aperture, the circle of confusion, the depth of field, the hyperfocal distance and the imaging law of the constant depth rendering, with analogy to a conventional optical imaging system. Based on the optical analysis of the virtual imaging sys-tem, we describe the relationship among the depth variation of the scene (depth of field), the constant depth(perfectly focused plane), the spacing of cameras (aperture) and the rendering resolution (circle of confusion). Specifically, the hyperfocal distance of the virtual optical system, as a key parameter for light field rendering, determines the relation-ship between the spacing of cameras and rendering reso-lution. Given the minimum and maximum depths of the scene, the optimal constant depth and the hyperfocal dis-tance are derived to achieve the best rendering quality or minimum number of images. The minimum number of im-ages required for anti-aliasing rendering (i.e., the rendering error is smaller than the circle of confusion) can be fur-ther reduced by segmenting the depth into multiple depth layers. A quantitative relationship between hyperfocal dis-tance, number of layers, and depth variation of the scene is described.