"Vision in the Real World: Finding, Attending and Recognizing Objects"

               Marten Bjorkman and Jan-Olof Eklundh
            Computer Vision and Active Perception Lab
          Royal Institute of Technology, Stockholm, Sweden

Abstract

We present a real-time stereo vision system able to locate and attend to
objects in cluttered realistic indoor scenes. The system is designed to
work as part of a larger robotic one, in particular for tasks that involve
the robot interacting with the environment. The system is built upon three
key components; an attentional process that delivers hypotheses to where an
object might be located, figure-ground segmentation that segregates attended
objects from its background and a recognition system that verifies whether
requested objects indeed have been found. Most parts of the system rely on
multiple cues for maximized robustness. In our study particular emphasis
has been placed on the binocular cues. To cope with the conflicting
requirements of a wide field of view and high resolutions, the system uses
two sets of cameras; a wide field set for attention and a foveal one for
recognition. In this presentation we will first give an overview of the
system and then concentrate on two different topics; foveated figure-ground
segmentation and multi-cue objects recognition.