A GPU-powered Computational Framework for Efficient 3D Model-based Vision

Abstract

We present a generic computational framework that exploits GPU processing to cope with the significant computational requirements of a class of model-based vision problems. We study the structure of this class of problems and map the involved processes to contemporary GPU architectures. The proposed framework has been validated through its application to various instances of the problem of model-based 3D hand tracking. We show that through the exploitation of this framework near real-time performance is achieved in problems that are prohibitively expensive to solve on CPU-only architectures. Additional experiments performed in various GPU architectures demonstrate the scalability of the approach and the distribution of the execution time among the involved processes.