Abstract:
One of the most fundamental computer vision problems is image
matching. We focus on the interesting case of two images of a
continuous surface. There is a significant body of literature on
defining parametric warps matching the two images. We show as a first
contribution that most of these warps lie in a general class, namely
Linear Basis Expansion with Multiple Parameter Sets warps. This class
of warps has a direct geometric interpretation: it is induced by a
deformable surface imaged by an affine camera. In practice, however,
the observed surface may not deform, but only undergo a rigid motion,
while the imaging geometry may lead to significant perspective
effects. This raises two fundamental questions: how can one constrain
the warps to comply with a rigid surface? Can the warps be extended so
as to deal with perspective projection? Our answers draw on two
contributions: the Single Parameter Set and Rational Basis Expansion
warps. Combining Linear/Rational Basis Expansion and Single/Multiple
Parameter Sets, we show that a full analogy exist with
Affine/Perspective projection and Rigid/Deformable surfaces. We derive
a general methodology that, from a standard Deformable Affine warp,
allows one to derive three new types of warps: Rigid Affine, Rigid
Perspective and Deformable Perspective.
Bio:
Adrien Bartoli is a permanent CNRS research scientist (Chargé de
Recherche) at the LASMEA laboratory in Clermont-Ferrand, France, since
October 2004 and a visiting professor at DIKU in Copenhagen, Denmark
for 2006-2009. Before that, he was a postdoctoral researcher at the
University of Oxford, UK, in the Visual Geometry Group, under the
supervision of Prof. Andrew Zisserman. He obtained his Habilitation
Degree (HDR - Habilitation à Diriger des Recherches) from Université
Blaise Pascal, Clermont II, in June 2008. He did his
PhD? in the
Perception group (ex MOVI), at INRIA Grenoble, France, under the
supervision of Prof. Peter Sturm and Prof. Radu Horaud and received
the 2004 INPG
PhD? thesis prize. Together with his
PhD? student Vincent
Gay-Bellile and his colleague Patrick Sayd, we received the best
student paper award at CORESA'07. Since September 2006, he is leading
the
ComSee? research team (with Thierry Chateau until early 2009). His
main research interests are in image registration and
Structure-From-Motion for rigid and non-rigid scenarios, and machine
learning, within the field of Computer Vision.
