View synthesis by the parallel use of GPU and CPU

摘要

We present an algorithm for efficient depth calculations and view synthesis. The main goal is the on-line generation of realistic interpolated views of a dynamic scene. The inputs are video-streams originating from two or more calibrated, static cameras. Efficiency is accomplished by the parallel use of the CPU and the GPU in a multi-threaded implementation. The input images are projected on a plane sweeping through 3D space, using the hardware accelerated transformations available on the GPU. A correlation measure is calculated simultaneously for all pixels on the plane and is compared at the different plane positions. A noisy 'virtual' view and a crude depth map result in very limited time. We apply a min-cut/max-flow algorithm on a graph, implemented on the CPU, to ameliorate this result by a global optimisation.