Synthetic Data Generation for Indoor Scene Understanding Using BIM

Visual facility inspections performed manually are tasks that can be automated. Segmentation of facility image data is one of the automated methods of identifying problems in facilities. However, the machine learning methodology that is mainly used to train the segmentation model requires a large amount of training dataset. Preparing training dataset accompanies laborious manual labeling. To address this issue, we present a new method for generating synthetic data that do not require manual labeling. The method is to create photograph-style images from the BIM images; a generative adversarial network called CycleGAN is used to enable style transfer between the two different domains.