Google took to its research blog today to tell the internet about Ground Truth, a technology based on deep learning that it uses to automatically update business listings on Google Maps. Essentially, Ground Truth uses geo-located Maps photos from the Google Street View team to populate new information about businesses, landmarks, street names, and other details. The algorithm is advanced enough to be able to work through samples that contain some blurry or scrambled photos, and even photos with multiple brands around the business name, as seen in the tire shop photo below. The best part is that the model behind the algorithm has been opened up to the general public, meaning that any developer can integrate the model to power their own autonomous information extraction functions.
The project got its roots back in 2008, when Google started using machine learning to identify faces and license plates in Maps so that it could blur them for the sake of privacy. This eventually led to Ground Truth's humble beginnings in 2014, when the algorithm began being used to read street numbers and help make addressing in Maps more accurate. From there, Google created the French Street Name Signs data set, consisting of over 1 million signs, to help train neural networks to read street names. Ground truth built upon that by reading other signs, including business storefronts, and integrating them into Maps. In its current form, Ground Truth scores a rather impressive 82.4-percent accuracy in tests using the French Street Name Signs data set.
Ground Truth had some important work done on it by Googlers who were interns at the time they did the job; an intern named Ian Goodfellow initially worked on the algorithm, while an intern named Zbigniew Wojna reworked the algorithm to allow Ground Truth to read and integrate street and business names. It was also the very first Google product to make use of Google's custom-made Tensor Processing Units, even before the consumer version of the similarly-named TensorFlow. Ground Truth is continuously being fed data at a rate of millions of photos per day, from Street View cameras worldwide, and will continue to improve its speed and success rate as it gets more and more on the job training.