LUNDI 12 NOVEMBRE, 14H, SALLE S16 (SALLE A CONFIRMER)

On the reconstruction of 3D shape from images

Carlos Hernández Esteban
Toshiba Research Europe Laboratory
http://mi.eng.cam.ac.uk/~ch394/

Abstract: Acquiring 3D shape from images is a classic problem in
Computer Vision occupying researchers for at least 20 years. Only
recently however have these ideas matured enough to provide highly
accurate results. In this talk we present three different algorithms
for 3D object modelling that are able to cope with objects of
increasing complexity: from the "simple" rigid objects,
to the more challenging untextured rigid objects and untextured
deformable objects.

For textured rigid objects, we propose a 3D modelling algorithm that
is able to efficiently fuse silhouettes and multi-view stereo for high
quality 3D object reconstruction. The algorithm is one of the top
performers in the recent evaluation of multi-view stereo algorithms by
Seitz et al.

If the object is not textured, we propose a novel multi-view
photometric stereo approach where the use of an uncalibrated distant
light source allows us to perform photometric stereo "in the
round" obtain a high resolution closed reconstruction. To our
knowledge, this is the first algorithm to reconstruct closed surfaces
using photometric stereo.

Finally, in order to deal with untextured deformable surfaces, we
propose a third algorithm using a setup with three coloured
lights. The algorithm can perform real-time reconstruction of
deforming untextured surfaces with constant chromaticity.