Leveraging deep learning segmentation techniques and connected component analysis to automate high-level cost estimates of facade retrofits using 2D images

Deep learning semantic segmentation techniques applied to 2D facade images hold a great promise in several domains that go far beyond model generation, mainly if the data used are front-parallel or orthonormal photographs. However, effective applications in the field of built heritage have not been adequately explored, largely due to the absence of multidisciplinary teams that include architecture professionals as early as the dataset creation stage. The aim of this research is to introduce a holistic view in order to demonstrate the practical usefulness of state-of-the-art segmentation models to automate high-level cost estimates of urbanscale residential building facade rehabilitations when combined with a connected component analysis. To achieve this, a scalable bottom-up approach is formulated in five simple phases, encompassing both data science and architecture expertise. This strategy seeks to improve the accuracy of analyses at early stages when limited information on constructions is available and there is a significant cost uncertainty, and therefore to optimise the strategies used by construction stakeholders involved in economic feasibility studiesand decision-making processes.