Geometry and Texture Synthesis

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

ABSTRACT 0204355 Greg Turk GA Tech Res Corp GIT Methods of data capture such as range scanning, still photography, video and motion capture have begun to alter the manner in which computer graphics modelling, rendering and animation is performed. We propose to widen the applicability of captured data for model creation by develop-ing methods for re-use of geometry and images. We plan to pursue two related research directions. The first of these is to improve methods of performing texture synthesis. We plan to address a num-ber of the shortcomings of current texture synthesis methods, including the loss or mis-registration of fine features, the lack of variation in synthetic textures, the tuning of the parameters of the algo-rithms and the lack of user control. Two of the ideas that we plan to draw upon for this work are image metrics for human visual perception and the evaluation of the quality of synthesized images. Surprisingly little attention has been paid to these issues with regard to texture synthesis. Our second goal is the adaptation of texture synthesis methods to the creation of 3D geometry. Specifically, we plan to address the problems of creating height fields, displacement maps, and arbitrary 3D volumetric geometry. There are many challenges to extending texture synthesis ideas into 3D, so we plan to approach this increasingly difficult list of representations in a methodical fashion. Of course, our work in texture synthesis will inform much of the geometry synthesis work. Geometry re-use should allow a user to create synthetic terrain with the locations of mountains, rivers and hills specified by the user. Natural erosion and other characteristic features will come not from costly simulations, but from geometry synthesis from examples. Examples of terrain data are easily found from the USGS, and other digital geometry can be captured through range scanning and merging techniques.