Building a web-scale image similarity search system

As the number of digital images is growing fast and Content-based Image Retrieval (CBIR) is gaining in popularity, CBIR systems should leap towards Web-scale datasets. In this paper, we report on our experience in building an experimental similarity search system on a test collection of more than 50 million images. The first big challenge we have been facing was obtaining a collection of images of this scale with the corresponding descriptive features. We have tackled the non-trivial process of image crawling and extraction of several MPEG-7 descriptors. The result of this effort is a test collection, the first of such scale, opened to the research community for experiments and comparisons. The second challenge was to develop indexing and searching mechanisms able to scale to the target size and to answer similarity queries in real-time. We have achieved this goal by creating sophisticated centralized and distributed structures based purely on the metric space model of data. We have joined them together which has resulted in an extremely flexible and scalable solution. In this paper, we study in detail the performance of this technology and its evolvement as the data volume grows by three orders of magnitude. The results of the experiments are very encouraging and promising for future applications.