Distributed Kd-Trees is a method for building image retrieval systems that can handle hundreds of millions of images. It is based on dividing the Kd-Tree into a “root subtree” that resides on a root machine, and several “leaf subtrees”, each residing on a leaf machine. The root machine handles incoming queries and farms out feature matching to an appropriate small subset of the leaf machines. Our implementation employs the MapReduce architecture to efficiently build and distribute the Kd-Tree for millions of images. It can run on thousands of machines, and provides orders of magnitude more throughput than the state-of-the-art, with better recognition performance. We show experiments with up to 100 million images running on 2048 machines, with run time of a fraction of a second for each query image. 1

To my family: my mother, my father, my wife and kids, and my sister. They have always helped and supported me.iv Acknowledgments First and foremost, all thanks are due to God, for giving me the strength and persistence to go through six years of graduate work. Second, I would like to thank my thesis adviser, Prof. Pietro Perona, for being an exceptional adviser, both academically and professionally. I learned a lot from his knowledge, insight, and guidance. Third, I would like to thank Prof. Yaser Abu-Mostafa. He was my initial co-adviser at Caltech. He gave me invaluable advice and patiently and graciously helped me refine and define my thesis topic. Fourth, I would like to thank Dr. Mario Munich for hosting me for an internship at Evolution Robotics, and for collaborating with me and providing me with advice during all my thesis work. Fifth, I would like to thank Dr. Jean-Yves Bouguet for hosting me for one internship at