Disertace

Ph.D.

ÚTIA AV ČR

ČVUT Praha

2006

1-133

Ing. Jiří Filip, Ph.D. (ÚTIA - Oddělení rozpoznávání obrazů)

texture, synthesis, BTF, MRF

Colour Rough Textures Modelling [file]

Constantly increasing graphics hardware computational power finally enables fast and realistic rendering of virtual reality models whose realisation was until recently impossible. Such realistic models require, among others, natural looking textures covering virtual objects of rendered scene. Applications of these advanced texture models in virtual reality systems now allow photorealistic material appearance approximation for such complex tasks as visual safety simulations or interior design in automotive/airspace industry or architecture.

For aim of such advanced applications a smooth textures lit by reflectance models alternatively combined with bump-mapping are not able to offer correct and realistic reproduction of material appearance. This is caused due to inherited complexity of many materials whose rough structure produces such visual effects as selfshadowing, masking, interreflection or subsurface scattering.

The one way how capture these material's attributes is using much more complex representation of a rough or 3D texture called Bidirectional Texture Function (BTF). BTF is a six dimensional function depending on view and illumination directions as well as on planar texture coordinates. This function is acquired as several thousands of images during varying light and camera positions. However, the huge size of measured BTF prevents it from using for any fast application so introduction of some fast compression, modelling and rendering method for BTF data is inevitable.

In this thesis we review BTF acquisition, modelling and rendering methods published so far, survey problems concerning BTF mapping and rendering implementation, surface height approximation and finally propose two novel BTF modelling approaches realised in several BTF modelling methods.

The first proposed approach introduces probabilistic BTF model based on Markov random field modelling of BTF subspaces. These subspaces are obtained using BTF segmentation and the regular material pattern is introduced into the model by means of surface height simulation.

The second one is based on polynomial extension of one-lobe Lafortune model computed in every pixel of original BTF measurements. The model is further extended by the parameters clustering to achieve higher compression ratio and the BTF sample enlargement by means of an image tiling technique.