One of the downsides of using simple devices for mapping the world is that the GPS accuracy is not always great, especially in cities with tall buildings. Since the start we have always wanted to correct this using image matching and we are now making progress in that area.
The technique is called 'Structure from Motion' (SfM) and means that you compute the relative camera positions and a 3D reconstruction of the environment using only the images.
We are now open sourcing our tools under the name OpenSfM and developing it in the open under a permissive BSD license. The project is intended to be a complete end-to-end easy-to-use SfM pipeline on top of OpenCV. We welcome all contributors, from industry and academia, to join the project. Driving this work inside Mapillary is Pau and Yubin.
Moving forward we are initially going to use this for improving the positioning and connection between Mapillary photos. Later, we are going to have an ever improving 3D reconstruction of every place on the planet too ;).
Here's an example from this Mapillary sequence. (takes a few seconds to load)
Hope you are as excited as we are in pushing this space forward in the open for everyone's benefit.