Rendering from Uncalibrated Stereo Images
Qian Chen and Gerard Medioni
We present an Image-Based Rendering approach which takes stereo image pairs as input. Unlike the pure image based approaches which restrict the ways the novel view is specified and the reconstruction based ones which require fully calibrated cameras, we compute a reasonable Euclidean approximation of the scene which allows to synthesize new views in a traditional way without significant visual distortion, while bypassing the requirement of internal camera calibration. This is achieved by the use of heuristics about the cameras themselves, and of constraints inherent in the stereo configuration. We convert the stereo matching problem into an extremal surface extraction problem, and provide a factorization-based algorithm for shape inference from correspondences. Results on real image sets are shown.