Inside Pastec, the images are represented using the visual word paradigm: http://en.wikipedia.org/wiki/Visual_Word During indexing, each image feature is assigned to the nearest visual word among a pre-trained set of 1 million. About 1000 visual words are extracted per image.