Segmentation, as the name suggests, consists of dividing an image into
"regions" that supposedly correspond to portions of the scene that
share some common property (for instance, belonging to the same
"object" in space). Unfortunately, segmentation from a single image is
intrinsically ill-defined, for there is no guarantee that the segmented
regions correspond to meaningful objects in space.
In this project, we advance the idea that segmentation can become a well-posed problem when multiple images of the same object (that satisfies suitable conditions on its radiance) are available. Then the "correct" segmentation corresponds uniquely to the three-dimensional shape of the object. Therefore, shape estimation can be posed as the problem of simultaneously segmenting several images of the same object, and viceversa. The practical consequence of this idea is an algorithm, based on a variational principle, that can estimate the three-dimensional shape of objects with smooth radiance or with very dense texture, conditions that render all existing parallax-based reconstruction algorithms useless.
This project is supported by NSF grants IIS-0208197 and IIS-9876145, ARO
grant DAAD19-99-1-0139, and Intel grant 8029.
© 2003 Hailin Jin, Stefano Soatto and Anthony J. Yezzi.
Please send your comments to firstname.lastname@example.org.
Last updated on January 8, 2004.
BACK to Hailin Jin's homepage