Multi-view computational scene modelling

点击翻译按钮获取中文摘要

In recent years, we have witnessed dramatic improvements in 3D reconstruction accuracy. With the advent of low cost depth sensors and more robust multi-view reconstruction algorithms, scene modelling techniques have become popularised and no longer limited to lab environments. However, despite their flexibility, existing techniques remain limited to scenes with simple surface reflectance properties (mostly diffuse). While more complex photometric approaches have been proposed to deal with such scenes, they tend to require expensive specialist equipment that cannot be deployed outside of the lab.This project will fill this gap by developing more general and versatile digitisation techniques that can be applied to natural scenes with complex reflectance captured using consumer hardware. Building on recent advances in computer vision and computational photography, the project will introduce novel acquisition hardware and algorithms to infer depth together with other important attributes such as surface normal and BRDF. To further push the limits of modelling accuracy, the research will also consider complex physical phenomena which degrade reconstruction accuracy, such as sub-surface scattering and inter-reflections, and will explore how algorithms can be made more resilient to these effects.The application of this research will be demonstrated in the context of the creative industries such as film, gaming and cultural heritage which all require the ability to create photo-realistic digital copies of assets. A particular emphasis will be placed on modelling natural or man-made scenes which have a complex surface reflectance (e.g. glossy materials). It is anticipated that this technology will find applications in many other areas of industry which require accurate modelling such as in robotics or the manufacturing industry.