Interactive view synthesis from compressed light fields

A light field is a collection of multi-view images which represent a 3D scene. Rendering from a light field provides a simple and efficient way to generate arbitrary new views of the scene as the viewing position and angle change, thus offering the experience of immersive viewing. The enormous amount of data required in a light field poses a key challenge in rendering. Tree-structured vector quantization (TSVQ) provides moderate compression ratio of around 24:1, which alleviates but does not solve the problem. Compression schemes based on video coding techniques exploit the data redundancy very effectively, but do not provide adequate random access for rendering. This paper describes a new compression scheme that supports interactive rendering directly from compressed light field data. The proposed algorithm provides a high compression ratio of as much as 10 times that of TSVQ, while only slowing down the rendering speed by a factor smaller than 2