When Google first introduced Street View, all it wanted to do was to provide panoramic views of the real world in its Maps. But then, as mobile devices and Google Maps shot up in popularity, the value of local search became obvious. And Google, as expected, became obsessed with providing accurate listings of local businesses across the world. In its possession were billions of high-res images captured by its Street View cars. But, who would analyze them? The task just didn’t seem humanly possible.
Enter, the wonders called machine learning and computer vision.
Today, after almost a decade of research, Google’s algorithm has become so perfect that over one-third of addresses globally have had their location improved through deep learning technology which recognizes text in its natural environment.
Google revealed in a blog post that this algorithm achieved an accuracy score of 84.2% on the very tricky French Street Name Signs dataset, which comprises of more than 1 million street names. This result is a radical improvement over earlier state-of-the-art systems. So much so, it has led to improvement in over 90% of the addresses in Brazil!
Google says, “Now, whenever a Street View car drives on a newly built road, our system can analyze the tens of thousands of images that would be captured, extract the street names and numbers, and properly create and locate the new addresses, automatically, on Google Maps.”
But, automatic address generation still doesn’t solve the purpose of local search – providing navigation to businesses by their name. So, the search engine giant proposed one more approach that would read store-front signs in Street View images. This model was trained extract ‘structured text’ from the imagery – meaning it would only keep that text which is relevant.
And here comes the cherry on the deep learning cake. Google has announced that it has opened the algorithm behind this model to the general public on GitHub through Tensorflow. This means, any developer can use it for automatic extraction of information from geo-located imagery. Go nuts, guys!