Hardware-Accelerated Dynamic Light Field Rendering

Abstract

We present a system capable of interactively displaying a dynamic scene from novel viewpoints by warping and blending images recorded from multiple synchronized video cameras. It is tuned for streamed data and achieves 20 frames per second on modern consumer-class hardware when rendering a 3D movie from an arbitrary eye point within the convex hull of the recording camera’s positions.The quality of the prediction largely depends on the accuracy of the disparity maps which are reconstructed off-line and provided together with the images. We generalize known algorithms for estimating disparities between two images to the case of multiple image streams, aiming at a minimization of warping artifacts and utilization of temporal coherence.