Building instances play a pivotal role in understanding population distribution and assessing vulnerability in the face of potential risks. Building instance segmentation is a valuable technique for identifying individual structures, however, complex urban environments pose great challenges, especially in the informal areas. In this study, we utilize a building instance segmentation method specifically designed to discern single buildings in both formal and informal settlements. Employing a SkipFuse-UResNet34 model, we generate building instances for Medellín, Colombia, resulting in a more comprehensive building mask compared to conventional official data sources. This enhanced mask serves as a vital tool for estimating the population at risk, enabling a thorough comparison with official data and addressing current spatial knowledge gaps.