How accurate should be disparities estimated for image-based rendering?

Abstract

Disparity estimation is the basis for the generation of a virtual view from a small set of real reference images. Much research on this area has been conducted by the computer vision community over the last decade. Currently, most problems involved in this technology are well understood and there exist several well-established algorithms to render virtual views from perspectively different images of the same scene. However, most work has been oriented towards high-accuracy disparity estimation to produce high-quality virtual images, often using sophisticated purpose-built hardware accelerators to achieve real-time results. Two disparity estimators with different complexity degrees are described and used to examine how much disparity inaccuracies influence image rendering quality. The objective of this study is to design software-based image synthesis in real-time on conventional PC platforms. Basically, this work looks at the opposite end of the cost-complexity curve by making very restrained demands on the disparity estimator. It is shown empirically that in many cases the effect of disparity accuracy on the quality of virtual views is almost imperceptible and that, for many applications requiring real-time processing, reasonably good results can be achieved with less computational cost.