Research overview

Most people find it easy to recognise objects based on their visual appearance. For a robot on the other hand, vision is a complex process that needs to be designed from scratch. The main focus of my research group is how to do this. Computer Vision for object recognition has been studied extensively in the past decade. If a recognition system is provided with a sufficient number of views of an object, it can be taught to recognise the object from arbitrary views. Computer Vision is however not yet on a par with human vision, and one possible reason for this is that Computer Vision usually is framed as a passive process, while the human visual system is inherently active. Object recognition systems are tested by providing them with an image, and the task is then to state what objects are present in the image. A human on the other hand, views objects with two eyes, can change gaze, and even head position in order to get a better view of the object.

The image above shows "Eddie the Embodied", the platform we are currently working on. Eddie has two cameras for eyes, and thus depth perception. His neck is a fast pan-tilt unit, that allows him to quickly shift gaze. Eddie can also speak, and thus report what he sees, and request objects to be moved or turned around. Later we will also add the ability to move sideways, and to stabilize gaze on moving targets.

You can find a movie clip of an early version of Eddie here.

On May 10, Swedish TV aired an interview where we presented Eddie. Click here to see it (in Swedish).

Under the headings Research projects and Publications in the left menu, you can find more information on what we are currently working on.