This is pretty impressive; a new algorithm that, when presented with a photograph with a hole cut out of it, searches a database of millions of other photographs, presents the user with a menu of similar-looking images to select from, and then composites elements of the chosen image to fill the hole seamlessly, producing an image which (in most cases) looks semantically coherent. Most impressively, it is entirely data-driven, and does not require any human-generated annotations of test data:
It uses mathematical properties of the images to make the match, and sometimes ends up serendipitously picking other images from the same location (because two photographs of, say, the Taj Mahal taken on a sunny mid-afternoon are likely to share similar properties).
Of course, it is possible to use such a tool creatively, replacing unwanted parts of an image with elements from a completely different scene, as the paper (PDF here shows:
