PhD Thesis : 2011

Modelling Appearance and Geometry from Images

Francho Melendez

The University of Manchester

A set of exemplars captured using our surface depth hallucination approach.


cquisition of realistic and relightable 3D models of large outdoor structures, such as buildings, requires the modelling of detailed geometry and visual appearance. Recovering these material characteristics can be very time consuming and needs specially dedicated equipment. Alternatively, surface detail can be conveyed by textures recovered from images, whose appearance is only valid under the originally photographed viewing and lighting conditions. Methods to easily capture locally detailed geometry, such as cracks in stone walls, and visual appearance require control of lighting conditions, which are usually restricted to small portions of surfaces captured at close range.This thesis investigates the acquisition of high-quality models from images, using simple photographic equipment and modest user intervention. The main focus of this investigation is on approximating detailed local depth information and visual appearance, obtained using a new image-based approach, and combining this with gross-scale 3D geometry. This is achieved by capturing these surface characteristics in small accessible regions and transferring them to the complete façade. This approach yields high-quality models, imparting the illusion of measured reflectance. In this thesis, we first present two novel algorithms for surface detail and visual appearance transfer, where these material properties are captured for small exemplars, using an image-based technique. Second, we develop an interactive solution to solve the problems of performing the transfer over both a large change in scale and to the different materials contained in a complete façade. Aiming to completely automate this process, a novel algorithm to differentiate between materials in the façade and associate them with the correct exemplars is introduced with promising results. Third, we present a new method for texture reconstruction from multiple images that optimises texture quality, by choosing the best view for every point and minimising seams. Material properties are transferred from the exemplars to the texture map, approximating reflectance and meso-structure. The combination of these techniques results in a complete working system capable of producing realistic relightable models of full building façades, containing high-resolution geometry and plausible visual appearance.




@phdthesis{phdthesis, author = {Melendez, Francho}, title = {Modelling Appearance and Geometry from Images}, school = {The University of Manchester}, year = 2011, }