Protecting Privacy in the World of Better Maps: How Collaboration Paves the Way
Having access to fresh street-level imagery is an important part of keeping maps and geodata updated. That’s why people and organizations upload millions of images to Mapillary every day—to understand their places of interest and improve their maps. By processing the imagery with computer vision, we are able to identify all the objects in the images and position them on the global map, making map updates easier than ever before.
We don’t want to position everything in an image on the map, though. At Mapillary, we build technology to allow anyone to understand as much as possible about every place that’s in an image. At the same time, we want to understand as little as possible about the people in it. That’s why we have built blurring technology for street-level imagery and incorporated it into our computer vision pipeline. Automatically blurring people’s faces and car license plates means we’re safeguarding privacy without compromising on fast and easy ways of updating maps.
We have known for quite some time that our blurring algorithms detect and blur more than 99% of all identifiable faces and license plates in street-level images. We’re not alone in doing automated privacy blurring through computer vision, though. Privacy is an important issue and technology can help protect it across many different use cases.
FPR: Percentage of incorrectly predicted pixels. Lower is better. All are evaluated on 2000 street-level images from Mapillary with manually annotated faces and license plates except for Google Street View, where the evaluations were done on Google’s internal datasets with their internal algorithms.
We recently looked at how other major players’ face detection algorithms available as public APIs perform on street-level images compared to Mapillary’s algorithms. As you can see in the table above, Mapillary tops the leaderboard. The results are astounding, but not surprising. What sets Mapillary apart from the rest is that our technology has been built solely for the purpose of street-level imagery. Our models have been trained and optimized for street scenes where there are small faces with varying poses, and the others have not. With this in mind, the results seem inevitable.
When looking at algorithms that are also optimized for street-level images, i.e. Google Street View, the result is not directly comparable due to the difference in the evaluation datasets. Nevertheless, our accuracy is on par with Google’s detection algorithms for both faces and license plates in street-level images.
The results of the benchmark mean that Mapillary has the best-performing algorithms for privacy blurring in street-level images, globally. On average, our technology misses just one out of 1,000 identifiable faces and license plates in images. And if you ever come across an image where the algorithms have missed a face or license plate, you can blur it yourself and your edits will help make the algorithms even better in the future.
These results are made possible through our community of individual mappers, cities, mapping companies, and everyone else who upload imagery to Mapillary to understand their places of interest and improve their maps. By uploading imagery to Mapillary, they don’t just get access to all the data they need to keep their maps updated, they also make sure that their imagery is anonymized and that privacy is protected across the board.
We often say that collaboration wins. Once again, it’s shown to be true. By working together, we’re all better off as a result. That’s why uploading imagery to Mapillary so that everyone can take part of the data is free of charge, and always will be. Upload your images by creating an account, or reach us on firstname.lastname@example.org.
/Yubin, VP Computer Vision