VASC Seminar Announcement
=========================
Date: Monday, 3/20/00
Time: 3:45
Place: NSH 3002
Speaker: Frank Dellaert
CMU Robotics Institute
http://www.cs.cmu.edu/~dellaert/
Title: Structure from Motion without Correspondence
Most if not all structure from motion approaches that work with sparse features
assume that the correspondence between features in different images is given,
or has been obtained in a separate pre-processing step, for example using robust
fundamental matrix estimation (The Oxford way :-).
In my talk I will present a method to recover 3D scene structure and camera motion
from multiple images *without* the need for correspondence information. Instead of
a separate preprocessing step that looks at pairs or triples of images, our method
works with all the images at the same time, and solves for structure, motion and
correspondence simultaneously.
To this end, we frame the problem as finding the maximum likelihood structure and
motion given only the 2D measurements, integrating over all possible assignments
of 3D features to 2D measurements. The approach is cast within the framework of
Expectation-Maximization, which leads to an intuitive iterative algorithm: at each
iteration a new structure from motion problem is solved, using as input a set of
'virtual measurements' obtained from a distribution over feature assignments. This
distribution can be efficiently obtained by Monte Carlo Markov Chain sampling.
The algorithm works well in practice, as will be demonstrated using results on
several real image sequences.
This is joint work with Steve Seitz, Chuck Thorpe, and Sebastian Thrun. The
technical report describing this work can be found at http://www.ri.cmu.edu/pubs/pub_3245.html.
A different version of this paper was accepted for CVPR 2000.