VIAMASS · Visual Recognition Made Super-Scalable

Hundreds of billions of images are hosted on the World Wide Web. There is a huge interest in mining information in large image collections based on visual content. Direct applications are the automatic organization of visual datasets, visual navigation, object recognition, and the traditional query-by-sample search.Although recent breakthroughs allow the search in millions of images on a single server with increasing quality, the accuracy of automatic recognition remains low compared to human’s visual analysis. I believe that significant progress is still achievable by a major shift in the paradigm underpinning the image representation: an image should be described with respect to the context provided by the image collection.The main objective of VIAMASS is to automatically discover visual links within a very large collection of images. These “visual hyper-links” will connect the objects across the images of the collection. This raises a major obstacle with respect to scalability: cross matching all the images is of quadratic complexity when performed with a brute-force approach. To this end, VIAMASS addresses issues at the frontier of the current state of the art in computer vision and signal processing: How to exploit the context provided by the collection to enrich the image representation? How to exploit and magnify recent signal processing and coding techniques to efficiently compare sets of vectors? How to automatically produce geometrical models of objects with little or no supervision? At the end, the ultimate challenge is to invent scalable solutions for the automatic discovery of visual links across images.My research program impacts the whole processing chain of visual search, from the description level to the mining algorithms that will break the complexity lock.