Computer Vision News - June 2016

when the image comes in, it is matched with all the nearby images (according to GPS coordinates), then they run Structure from Motion with everything nearby and they update the 3D model of that place and the positioning of that image. They also weigh images based on the quality of the GPS measurement, so that the more accurate images help position the less accurate ones . The rational for doing this is double: firstly, it improves the positioning of the images; secondly, the 3D reconstruction resulting from running Structure from Motion is actually the key to navigate between the photos. For, even if Mapillary supports 360 ° cameras, most of its imagery comes from regular phones: when you want to move around in space, it is very important to get correctly the special positioning of the photos, in order to transition between them in a nice way. Following this, the image is run through a recognition pipeline to detect the objects in the picture and all this data is finally made available. The higher the overlap, the better it is: the app takes high-res pictures at fast intervals, depending on the motion speed (driving, walking, etc.), to generate the highest possible overlap at high-pixel density . The image processing happening in the background is significantly complex and it is fully automated for hundreds of thousands of images every day. Solem says that building that infrastructure is probably the biggest technical challenge incurred by Mapillary’s engineers. One of the nicest features of the app is its ability to travel through time: if we see the same view at different times of “ Image processing is fully automated for hundreds of thousands of images per day ” 6 Computer Vision News Application