Journal of Computational Design and Engineering

To automatically generate data for training deep convolutional neural network models to segment building facades, researchers from Osaka University used a three-dimensional model and game engine to generate digital city twin synthetic training data. They found that a model trained on these data mixed with some real data was competitive with a model trained on real data alone, revealing the potential of digital twin data to improve accuracy and replace costly manually annotated real data.

Osaka University researchers created a diminished reality system for providing a future view of a building to be demolished. By implementing generative adversarial networks on a remote server, the team was able to stream real-time video that predicted what a landscape would look like after a building was demolished. This technology can help with urban renewal planning and stakeholder discussions.