Augment the road map data by the width
Nowadays there are many digital road maps available, even as free access. A good example is the OpenStreetMap (OSM) project, which contains a cartographic map of the road topology with good coverage over almost the full world. This is advantageous as it is freely available on the web and the quality and quantity of the annotations are growing over time, as more users contribute to the project. However, the map information is noisy and partially missing as for example most roads do not contain information about their width. In our research we propose to exploit aerial images in order to enhance open-source maps (e.g., with road geometry). This is not an easy task as despite decades of research, large-scale automatic road segmentation from aerial images remains an open problem.
This research was done in cooperation with the research group of Professor Raquel Urtasun at the Computer Science Department of the University of Toronto.
The problem of measuring the width of the road in an image can consider a semantic segmentation problem, where a binary label (road /non road) is assigned to each pixel of the image. This problem is well studied in computer vision, however it is not perfectly suitable for large areas. First, it has to reason about each pixel which would make it very computation intensive. Second, the true width might be hidden due to occlusion, e.g. by trees.
In contrast to traditional semantic segmentation problems, we propose to use OpenStreetMap (OSM) to formulate the problem as inference in a Markov random field (MRF) which is directly parameterized in terms of the centerline of each OSM road segment as well as its width. This parameterization enables very efficient computations and returns the same topology as OSM. In particular, we can segment the OSM roads of the whole world in only 1 day when using a small cluster of 10 computers.
For more details please read our paper “Enhancing Road Maps by Parsing Aerial Images Around theWorld”, Gellert Mattyus, Shenlong Wang, Sanja Fidler and Raquel Urtasun, International Conference on Computer Vision (ICCV), 2015. (A PDF of the paper is available on the right.)
Dataset, Image Features and Code
We provide the datasets used in the paper, the input features calculated for the images and the source code for the method. They can be found at downloads. If use the data or the code, cite our ICCV paper. Both the code and the data are intended for research and academic purposes, not commercial use.
An example from the Bavaria dataset. We also provide the annotated road masks.
An example for the output of our method. The yellow color shows the roads with their width extracted by our method.