Semantic segmentation of 3D point clouds
Autonomous vehicles require digital maps to navigate safely, smart cities requires knowledge of urban features to be managed appropriately, and digital twins require physical assets to be recognised before they can simulate predictive models. All those applications require a detailed representation and understanding of the spatial environment. The ability to create intelligent 3D models of the real world is a critical enabler for the reduction in cost and programme of major design activities. For instance, Highways England is one of the pioneers in developing parametric design solutions for complex national infrastructure.
SenSat captures high resolution images via drones with a ground sampling distance of ∼2.5 cm. Those images are then transformed into 3D point clouds, using techniques such as Structure from Motion (SfM). The ability to understand point clouds and 3D shapes is critical for supporting the growing market of parametric design and various applications.
The overall aim of this challenge is to investigate various semantic segmentation methods to discover how well they perform on outdoor large-scene urban area datasets. We also aim to find what are the limitations of the individual methods and how useful they are in practical problems such as the one considered within the challenge.
Data Study Group team. (2020, June 5). Data Study Group Final Report: SenSat. Zenodo. http://doi.org/10.5281/zenodo.3878499
Tarek Allam, UCL
Ondrej Bohdal, University of Edinburgh
Nanqing Dong, University of Oxford
Ivan Erofeev, University of Edinburgh
Qingyong Hu, University of Oxford
Julian Kuehnert, IPG Paris
Pauline Maury Laribière, SenSat
Luca Morreale, UCL
Mukharbek Organokov, University of Strasbourg
Chenfu Shi, University of Manchester
Wen Xiao, Newcastle University