Abstract: Recent advances in techniques for capturing large scale models of
urban environments, give rise to many novel applications which require
rapid and realistic 3D modelling. I will present an 3D reconstruction
approach utilizing properties of piecewise planarity and restricted
number of plane orientations to suppress the ambiguities causing
failures of standard dense stereo methods. I will describe how to formulate
this problem in MRF framework built on an image presegmented into superpixels and
demonstrate superior performance in problematic scenarios containing
many repetitive structures and no or low textured regions.
Using the same type of representation, I will briefly introduce some
on-going work on semantic parsing of urban areas using spatial
co-occurence of visual words and 3D geometry. I will show some
preliminary results on challenging environments with varying viewpoints and large number of categories appearing simultaneously.