Google is coming to The Computer Vision and Pattern Recognition Conference in Miami next week with a new cool research project. Google allows an internet user to provide an URL with an image of a historic site that she/he does not recognize and search a database of over 40 million geotagged photos (taken from Picasa, Panoromio and Wikitravel) to match that picture to verified landmarks. This early stage “landmark recognition engine” compare the image to a representative group of pictures of a landmark taken from similar perspectives. It sure might leverage Google computing power to perform the automated comparison super fast, we have heard that they have been testing GPU-based image recognition, so we wonder if GPUs are used in the back end.
Image recognition is not a new technology, but recognizing objects is probably more difficult than face recognition. The geotagging might help the process a lot because Google is probably building data structures based on spatial partitioning. That reduces the number of comparisons needed to return a search result.
European researchers have developed a similar application named MOBVIS system that recognize individual buildings in a photo shot from a camera-phone, no details is given about the technique in this article from Science Daily, and we do not know whether automatic geotagging available in the phone is used or not. If automatic geotagging is used (probable), it is making things way easier for MOBVIS system than what the Google engine is doing.