Section: Software

Large-scale image search

Participants : Matthijs Douze [correspondant] , Mohamed Ayari, Cordelia Schmid.

LEAR's image search demonstration was extended to 100M images. The image dataset was provided by Exalead. Search at this scale is possible due to the Fisher vector representation and the pqcodes software. The search time on a single core is about 250 ms.

In collaboration with Hervé Jégou, from the INRIA Texmex team, we stabilized and improved the pqcodes software package. The software was extended to implement matrix multiplications in the PQ-compressed domain. A non-exclusive license on pqcodes was sold to Technicolor. Another agreement is under negotiation with Morpho (a company owned by Safran).

LEAR's implementation of the Fisher descriptor was improved in several ways. A new method to train the GMM was developed and the computation time of second-order derivatives (w.r.t. $\sigma$) was significantly reduced. Furthermore, the extraction of dense SIFT descriptors was improved in quality and speed.