A hole in a picture filled from other photographs
In August 2007 at SIGGRAPH James Hays and Alexei Efros of Carnegie Mellon filled a cut-out part of a photograph with pieces of other photographs. First the 200 most similar scenes were sought among 2.3 million Flickr images, then the best place in each, and the insert was stitched in without a seam.
Why it matters
Completion methods had taken material only from the same picture, so they could extend grass or a wall but not paint in something absent from the frame. Here data replaced the model: with ten thousand images scene matching was useless, with two million there was, in the authors’ words, a qualitative leap. The 2009 essay "The Unreasonable Effectiveness of Data" would cite the example.
ACM Transactions on Graphics 26(3), SIGGRAPH 2007. 2.3 million unique images, 396 GB of JPEG, from thirty Flickr groups and keyword searches; scenes are compared with the gist descriptor, and placement within 80 pixels around the hole. Over an hour per image on one processor, about 5 minutes on fifteen. In a study with 20 participants and unlimited time, the results were taken for real photographs 37% of the time against 10% for the method of Criminisi et al.; real photographs 87%.