high performance computing on graphics processing units: hgpu.org

Capturing, recreating and representing a high fidelity virtual representation of the dynamic human form has long been a target for a diverse range of applications including tele-presence, games, film and TV special effects. The complexity of the challenge, to achieve a lifelike, faithful and believable representation, is such that a wide range of techniques and approaches have been developed. These are both due to research lead curiosity and requirements to address specific objective for particular problems. This work starts from a novel standpoint: that the processes of surfacing, tessellation and texturing, commonly used in 3D reconstruction, are computationally expensive and un-necessary. This work argues that by integrating the reconstruction and rendering processes into a single process that is aligned with the architecture of modern graphics hardware, a lightweight component solution can be achieved that is suitable for application on the end user systems within the many application domains. In order to achieve this aim the research undertaken seeks to both define an appropriate technique and develop detailed understanding of the reconstruction process pipeline and impacting factors. This is achieved through a complementary investigation of the tools and frameworks that are necessary to support iterative development of the approach with reliable, repeatable objective assessment. This reasons that by understanding the nature of the capture, reconstruction and presentation pipeline and by objective evaluation of the emerging reconstruction techniques this research will define an approach for 3D video based reconstruction that effectively utilises the processing potential of a single system to deliver acceptable levels of performance (speed) and fidelity (visual quality) for a componentised, multi-purpose 3D reconstruction and rendering solution. This thesis describes the research that has driven the evolution of technique and documents the iterations made. It presents a novel framework for experimentation and evaluation of the techniques and demonstrates how the use of these tools has enabled both rapid prototyping of approach and objective evaluation of improvement. The work concludes with a review of the approach taken and identifies approaches for evaluation of performance (speed) and fidelity (visual quality) that enable both repeatable experimentation within the research pipeline and reliable comparison of the end-to-end process against other techniques.