Object contour is arguably one of the most robust visual features. Even on transparent, textureless, or extremely shiny objects, contour extraction boils down to simple thresholding or edge detection, if the background and illumination can be controlled. The amount of information, which is embedded in object contours, is best demonstrated by the human visual system. A human is able to recognize objects from their binary images and gets a good idea of object shape and motion from a temporal sequence of binary images. In contrast to a human, a computer vision system still has a hard time in efficiently exploiting contour information. A large amount of scientific effort has been put into visual hull reconstruction, i.e. the problem of shape retrieval from contours, robust shape matching and contour based pose estimation, but the following basic questions are still open:
How are partial and closed contours efficiently represented and compared?
How can the space of possible contours from all views of a known object be efficiently modelled?
How can the space of possible contours within object categories be efficiently modelled?
In the proposed project, we will focus on these questions. In the context of pose estimation, we will examine how measured object contours can be efficiently matched against a known 3D model. Pose estimation techniques currently employ local, iterative methods to accurately align object shape with a contour. These methods are likely to get stuck in local minima, and are not able to resolve symmetries. We believe, that it is possible to discretize the space of possible views of an object and organize it in an efficient manner. An efficient matching technique against this contour space would result in a globally optimal match.
We will further examine, how multiple, simultaneous views of an object can be used to resolve symmetries and other ambiguities.