In a nutshell at this moment it is classification tool, but not based on pre-trained model. So, when it looks for feature index does it pick the first available (or) it iterates for best available? good luck!
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.