Noise Removal in Building Documentation by Generative Adversarial Networks

In our day, it is possible to document existing buildings and transfer to digital medium in 3D owing to the advancements in the technology. In the course of this transfer, temporary, permanent or seasonal obstacles may reside between buildings and data capturing devices regardless of the method of data collection and as a result preventing the precise and correct documentation of the buildings. In the scope of this project, it is aimed to obtain high precision 3D models reflecting the real case of buildings with the utilization of deep learning based models to clear these obstacles which can be regarded as noise in the data autonomously. In this context, a method which the noise in building photos are autonomously detected and noise regions are inpainted with respect to the building for 3d reconstruction software from multiple photos is proposed. Semantic segmentation and generative adversarial networks based inpainting models are utilized in an end-to-end manner. The testing of the proposed method is conducted in both indoor spaces and building facades and the results revealed an increase in the precision of the 3D models with respect to the models obtained from unprocessed photos.