Abstract : Traditionally, images distortions have been regarded as a flaw that needed to be corrected. Much effort has been focused on “undistortion” methods to remedy this shortcoming. However, we see distortions as encoding crucial information about scene structure. In this paper we describe how scene depth is encoded in the distortions of images acquired with non-single viewpoint cameras or by mosaic construction. We present our framework to exploit these distortions for 3D euclidean reconstruction of scene features from a single image or view. We present methods specifically designed for features such as straight lines, circles or conics and show how it applies to general planar curves as well. Rigorous experimentation using simulations and synthetic images from catadioptric sensors and mosaics are presented.
Type de document :
Communication dans un congrès
The 8th Workshop on Omnidirectional Vision, Camera Networks and Non-classical Cameras - OMNIVIS, Oct 2008, Marseille, France. 2008
https://hal.inria.fr/inria-00325383
Contributeur : Peter Sturm
<>
Soumis le : lundi 29 septembre 2008 - 10:39:44
Dernière modification le : lundi 8 octobre 2018 - 20:46:07
Document(s) archivé(s) le : jeudi 3 juin 2010 - 22:06:48
Rahul Swaminathan, Ao Wu, Haoyuan Dong. Depth From Distortions. The 8th Workshop on Omnidirectional Vision, Camera Networks and Non-classical Cameras - OMNIVIS, Oct 2008, Marseille, France. 2008. 〈inria-00325383〉
