SpaceNet Challenge 7 Results and Winning Models Released
The results of SpaceNet Challenge 7 have been released along with the winning models. The challenge was initially announced in July 2020 and saw a large number of submissions. The challenge was focused on the SpaceNet7 dataset, which is made up of two years worth of monthly images for 100 unique regions around the world, coupled with annotations marking the buildings present in each image. The images were captured by Planet satellites and manually annotated by a team at SpaceNet. The goal of the challenge was to develop a model capable of detecting buildings in images, and mapping individual buildings across timesteps. The ability to automatically detect urban change and development is useful in many domains, from assessing the risk posed by natural disasters, to producing records for land registries.
The 5 top scoring models all used various combinations of convolutional neural networks to first extract the building footprints from images, and then made use of a variety of methods to track individual buildings through time. Using the performance metric defined in the challenge the initial baseline model achieved a score of 17 of a possible 100. The winning model developed by four Baidu engineers achieved a score of 41 which is a very significant improvement. A full analysis of the winning models can be found in this paper.