XMedia Project
X-Media addresses the issue of knowledge management in complex distributed environments. It studies, develops and implements large scale methodologies and techniques for knowledge management able to support sharing and reuse of knowledge that is distributed in different media (images, documents and data) and repositories (data bases, knowledge bases, document repositories, etc.). The project started in March 2006 and will last for 4 years. It has a budget topping Euro 13.6M, (9.9M from the EU). 15 partners are involved from UK, Germany, Italy, France, Slovenia, Greece and Norway.
ISE was (partly) funded by the X-Media project, sponsored by the European Commission as part of the Information Society Technologies (IST) programme under EC grant number IST-FP6-026978.
Developments of ISE in the context of the XMedia project led to the enrichment of ISE with new feature descriptors (MPEG-7 feature descriptors, statistical texture descriptors) and new algorithms (clustering kmeans++, dbscan, mean shift).
External Tools based on GIRL
In the context of the XMedia project, external tools based on GIRL/ISE libraries have been developed
We have proposed a content based image indexing method based on a Bag-of-Regions approach. The content of the image is expressed using visual words computed by the clustering of image regions obtained by a unified segmentation algorithm and using several regional descriptors. This work is still ongoing.
Another example is the joint segmentation and recognition of objects in still images. Our approach relies on image over-segmentation and supervised training based on SVM. Regional descriptors are used to capture the color and texture content of the segmented regions, and a supervised SVM training algorithm achieves the mapping between the low level regional descriptors and the associated high level concepts. The region classification obtained is further enhanced by taking into account the local relationships between regions estimated on the training based, using an iterative refinement algorithm of relaxation labeling. The videos below demonstrate the approach on the Microsoft Research 21-object class database. The original image is presented along with its segmentation mask. Each color represents a specific concept class. The videos starts by showing the initial region classification obtained using the SVM classifier, and illustrate the iterations of the relaxation algorithm until convergence.
Original
Region Classification using SVM
Relaxation Labelling