Recovering Complete Scene Structure from a Sequence of Images
Steve Seitz
Computer Sciences Department, UW-Madison
2:30pm, Friday May 12 in 2310 CS
Although the problem of computing scene structure from images has been well
studied, virtually no attention has been given to assessing how much of the
scene can be reconstructed. The result is that current techniques are capable
of accurately reconstructing a relatively small proportion of the visible
scene. This talk presents a novel approach to image-based structure recovery
that guarantees the completeness of the recovered scene, ensuring that every
continuously visible scene feature is reconstructed. The algorithm relies
on four or more reference features that have been tracked through the image
sequence. The motion of the reference features is used to establish
constraints that simplify the recovery of the rest of the visible scene.
In addition, the technique is immune to outliers and feature drift--two
weaknesses of existing structure-from-motion techniques. Results of the
algorithm running on real image sequences will be presented.