George (Jiri) Matas

photo photo photo Contact Information
 Phone: +44 1483 876038
        +420 2   24357212
 Fax:   +44 1483 879554
        +420 2   24357385

I am mainly interested in both theory and applications of object recognition. Currently, most of my time is spent on the problems of colour-based image retrieval, face verification and on recognition using invariant pixel sets. In the last couple of years I have published on colour-based recognition and segmentation, face authentication, Hough Transform, lip tracking, active vision , combination of evidence and other topics not so easily classified (clearly I need more focus!). The complete list of publications has about 90 items. Lets look at some of the results more closely, mainly through a couple of demos:

With my PhD. student D. Koubaroulis we proposed a new representation of object colour structure for retrieval and recognition , called the Multimodal Neighbourhood Signature (MNS). Local colour structure is represented by illumination invariant features computed from image neighbourhoods with multimodal colour density function. For details, look at the MNS image retrival demo page (it includes pointers to publications on the method).

The Invariant Pixel Set method for recognition tries to find regions in the image that can be separated from the background in a viewpoint independent way. Convex sets defined by Harris corners could serve as and example as described in the BMVC 2000 paper. Look at a demo.

Since 1997 I have been working on Face Authentication (face identity verification) for applications in biometrics, first within the M2VTS project and currently in the 5th framework BANCA project. With K.Jonsson we developed a Prototype System for Face Verification. I contributed to reseach on the use of Linear Discriminant Analysis for face recognition and proposed a novel method for robust registration of images. For more details, look in my publications on face related topic. Some time ago I supervised a project on real-time blink detection. The sequences with blinks marked by white boxes are fun to watch. Look at the woman and man sequence (there is one false positive on the man sequence).

In 1996-1997 I worked with Ulises Ramos on a chromaticity driven Lip Tracker, testing it on the M2VTS database. The M2VTS tracking demo page shows the results and includes pointer to our publications on the topic. Before completing his PhD thesis, Ulises tested a robustified method on the 2760 sequences of the multimodal XM2VTS database with 98.5% success rate! All the tracked contours are available on-line, see the XM2VTS tracking demo page


Source code for the libraries is available on ftp as comressed tar files from this directory., $Date: 2002/08/30 12:53:48 $