Scene reconstruction from multiple views using a locally planar model

Image based rendering techniques allow reconstructions of a scene from an arbitrary viewpoint, given views of the scene. In this paper a locally planar representation of the scene structure is presented, where each block of pixels in the input images is assumed to correspond to some patch of an object's surface. A multi-resolution particle filter method of estimating the position and orientation of these patches is described. This method offers fast computation without sacrificing accuracy. The estimated patches often include a significant amount of redundant information. We present an algorithm for identifying this redundancy and producing a compressed data set. Finally we introduce a simple, fast, algorithm for generating reconstructions of the scene. Results on both synthetic and real data sets are presented to demonstrate the effectiveness of our techniques.