Large-scale image indexing
Participants : Matthijs Douze, Hervé Jégou, Benoit Mordelet, Cordelia Schmid.
Our large-scale image indexing software has been improved in 2009 as follows:
the point-based scoring method now includes the burstiness measure 
several algorithms that index global descriptors have been added. The first one builds an inverted file from GIST descriptors  , the second one indexes several "MiniBofs"  .
An efficient high-dimensional nearest-neighbor method has been added: the product quantizer  .
A client-server version of the video search engine is now included. It is aimed at application scenarios where the server should not have access to the unprocessed video data.
The on-line demonstrator now indexes 10 million images (2 million in the 2008). It can tested at http://bigimbaz.inrialpes.fr .
Version 2.4 of LEAR's image search engine, Bigimbaz , has been registered at the “Agence pour la Protection des Programmes”, under IDDN.FR.001.510004.001.S.A.2008.000.21000. It has been transferred for research purposes only to Stanford University, San Diego University, and the California Institute of Technology.
We have also adapted and tested the software for evaluation by the i-Dash European project (http://www.i-dash.eu/ ). This project aims at producing a platform that the police can use in their investigations which involve large quantities of child abuse video material. Our video indexing system can be used to recognize strongly distorted and low-quality video clips. It has been evaluated and shortlisted by TNO (http://www.tno.nl/ ) to detect known instances of illegal videos.