When most of us travel, we like to visit famous landmarks in the area we’re touring. It’s easy to find famous landmarks, like the Eiffel Tower or the Great Pyramid, but can be difficult to find lesser known sites. And for armchair travelers, it can take time to search for images of landmarks even with powerful image search tools that use GPS tagged photos or offer virtual tours.
A group of researchers from Google and the National University of Singapore have helped solve this problem, by building a “web-scale landmark recognition engine.” They’re presenting a report on their work today at the Computer Vision and Pattern Recognition (CVPR) conference in Miami, Florida.
This is similar to the work UCSD researchers did using Google data centers a couple of years ago. That work, called Supervised Multiclass Labeling (SML), involved systematically training a computer to recognize statistically similar objects, and tagging the images with labels that described image characteristics. For more on that project, see Teaching Google To See Images.
By contrast, the landmark recognition engine first had to gather the names of landmarks, and then associate those names with photos of specific landmarks. The researchers used clustering techniques to find similar images, filter out “noisy” images, as well as eliminating false positives (images that look similar but depict two entirely different things).
The researchers compiled a comprehensive list of landmarks from more than 20 million GPS-tagged photos and images located on online tour guide web pages. Armed with this list, candidate images for each landmark were obtained from photo sharing websites and by querying an image search engine. The resulting landmark recognition engine successfully identifed images of more than 50,000 landmarks from all over the world with 80% accuracy.
Unfortunately, Google has no immediate plans to make the new landmark search engine public, or incorporate it into image search. But you can read more about it in this official Google blog post, A new landmark in computer vision, or if you really want to dive into the technology, read the paper the researchers are presenting: Tour the World: building a web-scale landmark recognition engine (PDF).