CLOSE

Reseach Result Generating synthetic datasets from a city digital twin

Osaka University researchers developed a novel framework that can automatically produce synthetic datasets from a city digital twin (CDT). The method automatically generates instance-annotated synthetic data of building facades using the CDT. When the deep neural network was trained on the CDT synthetic data, the extraction accuracy of instance segmentation of building facades was improved over that of training on virtual data with no corresponding buildings in the real world. The performance of instance segmentation was improved by extending the proposed CDT synthetic data and mixing it with real-world data. The proposed automated system significantly reduces the cost of dataset creation.

Instance segmentation of building facades trained on CDT synthetic data and real world data
Overview of the proposed method
Comparison of CDT (top) and real-world street-view (bottom) from the same perspective (Koto city, Tokyo)

Published paper

Press Release(ResOU)

(YouTube) Generating synthetic datasets from a city digital twin