Abstract

Abstract: The first half of this talk will describe scalable methods for performing particular object search via visual queries in very large datasets of unordered images. We show how the combination of large visual vocabularies, fast spatial verification, query expansion and soft assignment can dramatically boost the precision and recall of object retrieval to give a fast and accurate system.

The second half will explore object mining in very large collections, where the aim is to automatically discover and group images containing the same object. We use object search to construct a matching graph over the collection where the strength of the edges depends on the spatial consistency between the image nodes. Standard clustering techniques are then used to mine the objects.

These methods will be demonstrated on several large datasets of Flickr images.

Joint work with Ondrej Chum, Michael Isard, James Philbin and Josef Sivic.

Biography: Andrew Zisserman is a Professor in the Department of Engineering Science, University of Oxford. He graduated from the University of Cambridge with a degree in Theoretical Physics, and for the last 20 years has carried out research in the area of Computer Vision. He has co-authored and co-edited several books on this area. The most recent, “Multiple View Geometry in Computer Vision” (with Richard Hartley), has now been published as a second edition in paperback and also translated into Chinese. He has been a program co-chair and general co-chair of ICCV . Software from his research group was marketed by the spin out company 2d3 (www.2d3.com) as a camera tracker for the special effects industry. This was awarded a Technical Emmy in 2002. He has been awarded the IEEE Marr Prize three times.