Russian Federation - 2016
For groundbreaking data-driven approaches to computer graphics and computer vision.
The world is swimming in images, and we need tools to help us organize, understand, and display this ocean of data. Efros is a pioneer in this effort, developing data-driven methods to both understand and create images.
Efros introduced methods for automatic copying and pasting from other images. He and his collaborators developed elegant algorithms to "steal" and assemble small patches from other images, achieving ground-breaking realism in synthesizing large regions of texture from small examples of the texture. He later extended that approach to large regions of images, showing that with a large enough database, good matches could almost always be found to remove and replace large parts of any image.
Building on the same insight of manipulating many small image examples, Efros and his students have been able to reconstruct the shape and surface orientations of complicated real-world objects from single images, to date photos by their appearance, and to identify characteristic image qualities, to let us address the question, "What makes Paris look like Paris?"
As his methods are helping us organize and create imagery in our fast-growing photo world, Efros is also helping us navigate the biases in dataset collection that affect the accuracy and fairness of data-driven algorithms. Through his pioneering data-driven approaches, Efros has had immense impact in computer graphics and computer vision.