High volume offline image recognition

Abstract

We show a prototype of an offline image recognition engine, running on a tablet with Intel®AtomTM processor, searching within less than 250ms through thousands (5000+) of images. Moreover, the prototype still offers the advanced capabilities of recognising real world 3D objects, until now reserved only for cloud solutions. Until now image search within large collections of images could be performed only in the cloud, requiring mobile devices to have Internet connectivity. However, for many use cases the connectivity requirement is impractical, e.g. many museums have no network coverage, or do not want their visitors incurring expensive roaming charges. Existing commercial solutions are very limited in terms of searched collections sizes, often imposing a maximum limit of <100 reference images. Moreover, adding images typically affects the recognition speed and increases RAM requirements.