Improving 2D Construction Plans with Cycle-Consistent Generative Adversarial Networks

Digital methods such as building information modeling (BIM) offer significant potential in the operation phase of a building. A prerequisite for these methods is the availability of digital as-built models of the existing buildings that are not given for many of them. This makes a retro-digitization, i.e., the generation of a digital representation of an already existing building, necessary. Thereby, computer vision (CV) and machine learning (ML) are essential technologies to extract the information from the data sources of a building. An important data source is 2D construction plans from which the initially planned geometry of existing buildings can be extracted. The extraction requires a highly reliable detection of lines and texts. However, the performance of the CV and ML methods is highly dependent on the quality of the data. Anomalies like discolorations, stains, and fold lines on 2D plans can negatively affect the detection. This is especially the case for old hand-drawn paper plans that are often the only data source available for old buildings. To integrate old 2D plans better in the detection process by CV and ML, their quality recovery is necessary. To achieve this, we propose using the cycle-consistent generative adversarial network (CycleGAN) that enables style transformation with unpaired data. Hereby transformation means the removal of the stated anomalies. Our results show that both text and edge detection methods perform better on improved 2D plans than on the original 2D plans with anomalies.