Image and Video
Topology Extraction of Roads and Buildings from Satellite Images
As the increasing popularity of navigation needs, people have become more reliant on navigation maps than before, especially on three services, POI searching (where to go), localization (where am I) and navigation (how to go). However, all these services heavily depend on map data completeness and accuracy. Data incompleteness and inaccuracy not only hurt end user’s experience but also can cause waste of time and inconvenience to vehicles and drivers. To produce map data, many different data sources have been exploited, such as GPS data, user feedback and image capture. Among them, satellite images enjoy the advantages of wide coverage of real world and overhead views, and they have already acted as an indispensable data source to extract road and building information for map data production.
Although there has been a lot of research on road and building information extraction from satellite images, e.g. road semantic segmentation and building instance segmentation, there is still a large gap from segmentation results to automatic map data production. To reduce the gap, techniques not only should work well on a wide range of different resolution images, but also be able to extract more precise representation, e.g. topology of the roads and buildings. This project aims to move one step forward to reduce this gap and achieve automatic map data production.
- End to end generalizable approach to different resolution satellite images
- Road vectorization and topology representation with error smaller than 2m
- Building footprint polygonal representation with IOU > 0.75
Related Research Topics
- Semantic segmentation from satellite images
- Object detection and instance segmentation from satellite images
- Polygonal representations of buildings
- Topology representations of roads
Suggested Collaboration Method
ARF (Alibaba Research Fellowship), three months ~ one year onsite.