2017 Year in Review
2017 will go down in Mapillary history as the year when we celebrated reaching 200 million images. Thanks to the collaboration of our whole community, including customers and partners, we now have 227 million images covering 3.9 million km, with 18 billion objects detected. Out of these, more than half came in 2017!
We are happy about our continued cooperation with OpenStreetMap, as well as extending the partnership with HERE and the Map Creator community. Collaborating and sharing data is the only sensible way forward and in 2017 we saw remarkable initiatives from cities (such as Amsterdam), countries (such as Lithuania), and companies alike (such as Microsoft). Individual volunteers, local organizations, and NGOs are doing an extraordinary job—from contributing 8 million images within a year like jbthemilker to organizing a Mapillary conference like our Bangladeshi community to covering their whole capital region in just a month like Bike Ottawa.
Throughout the year, we added more and better map data to the platform. You can visualize image detections, where our AI algorithms classify each pixel in an image. We added a new type of map feature derived from these detections—19 classes of objects are now extracted as line features. And for traffic signs, which have been available as point features already for a while, we added support for 500 more signs in 60 countries. This means we now recognize more than 900 different ones all over the world.
To increase the accuracy of the map data we provide, we put a lot of focus on the Human in the Loop approach, and released a fun game for verifying traffic sign detections as well as the tagging tool. This helps organize imagery and is also an important step on the way to make custom detections available. Verification tasks, as well as capture and map editing tasks, can be set up in the new Mapillary Tasker, which connects contributors with projects (supply with demand). For capture tasks, there is now the #CompletetheMap framework for targeted collection of imagery.
Our API went from v2 to v3, offering new possibilities for accessing Mapillary data. And MapillaryJS got support for map object editing, which has also been weaved into Esri tools. We’ve seen some great integrations built this year, using our API, MapillaryJS, and vector tiles. Some examples are the OsmAnd Mapillary plugin, Mapillary imagery in Eniro’s location services, and the first street-level views of Sarajevo.
In the spring, we released the Mapillary Vistas Dataset—25,000 pixel-wise annotated images used for training AI algorithms. This was of interest to companies such as Toyota, Lyft, Daimler, FiveAI, Volvo, and Zenuity (thanks for sponsoring the dataset!). Our own Research Team has achieved noteworthy results when using this data. By developing a novel approach to training neural networks, they are setting new benchmarks and changing the way memory usage can be optimized in modern deep learning architectures.
We want to say a big thanks to everyone who’s been working with us and wish you a happy and successful 2018!