Computer Vision News - June 2016

For example, many cities buy Street View to make an inventory of things that are in the streets landscape, to monitor their signage or the road’s quality, to assign building permits or design city planning: each time they request a capture of the city, Solem explains, it costs hundreds of thousands of dollars . With this solution, they can take the pictures themselves anytime and even ask citizens to contribute by taking pics. Mapillary’s app creates the vision for them , which means enormous cost savings via a system that is immediate and many times faster than available alternatives: this is where Solem sees the business. Right now, when ordering a company to come and scan your city with Street View cameras, it takes months to place that order and another few months to publish the images, whereas with Mapillary, anyone can take a picture anywhere and it will be available only a few minutes after. Mapillary does things differently from other companies operating in the same business space: without owning a fleet of cars, drivers and vehicles to drive on and capture thing; without counting on hundreds of people to cure and edit the data; Mapillary has set up a small team around which everything is automated, at least on the imaging side. They receive daily hundreds of thousands of images , which need to be processed according to a number of steps: faces and license plates are blurred and then pictures are published so that people can see them quickly online, as expected by the uploaders. Following that, images are sent to a pipeline that does an incremental type of Structure from Motion . Not having control of when the photographs are coming, from which camera they are taken and from which angle, Mapillary has a challenging problem compared to many others that do a controlled capture: the latter take all the images back to the office, process them in batch and upload the result, while Mapillary must be able to reconstruct on the fly in 3D every image that comes in. So they invented a clever solution: Computer Vision News Application 5