JCSE, vol. 2, no. 1, pp.98-108, March, 2008

DOI:

Terrain Geometry from Monocular Image Sequences

Alexander McKenzie Eugene Vendrovsky Junyong Noh
Caltech|Rhythm Hues Studios|KAIST

Abstract: Terrain reconstruction from images is an ill-posed, yet commonly desired Structure fromMotion task when compositing visual effects into live-action photography. These surfaces arerequired for choreography of a scene, casting physically accurate shadows of CG elements, andocclusions. We present a novel framework for generating the geometry of landscapes fromextremely noisy point cloud datasets obtained via limited resolution techniques, particularlyoptical flow based vision algorithms applied to live-action video plates. Our contribution is anew statistical approach to remove erroneous tracks (??痍펤liers?? by employing a uniquecombination of well established techniques?including Gaussian Mixture Models (GMMs) forrobust parameter estimation and Radial Basis Functions (RBFs) for scattered data inter-polation?to exploit the natural constraints of this problem. Our algorithm offsets the trem-endously laborious task of modeling these landsc

Keyword: No keyword

Full Paper:   238 Downloads, 3972 View