Séminaire ETIS : Hervé Jégou
Titre du séminaire et orateur
Approximate search as a source coding problem, with application to large scale image retrieval.
Hervé Jégou, INRIA Rennes.
Date et lieu
Mardi 4 décembre 2012.
ENSEA, salle du Conseil.
Image recognition, which is used in many applications such as copy detection or location recognition, requires to handle and search into large databases of descriptors, typically in the order of billions of vectors. This raises an efficiency problem, but also the problem of memory resources.
In this talk, I will first show that the search problem can be cast into a source coding framework, where the database vectors are approximated by product quantization. The Euclidean distance between a query vector and a database vector is estimated in an asymmetric manner based on the quantized database descriptors, thanks to the properties of the product quantizer. The method is advantageously combined with an inverted file to avoid exhaustive search, and used either for local or global descriptors.
I will finally consider of the interest of this kind of technique in a context of large-scale image search and classification.