Exploración de las capacidades del modelo Segment Anything Model v2 (SAM 2) para la segmentación de fachadas en el entorno urbano de Bogotá a partir de imágenes de Google Street View

Abstract

This undergraduate project addresses the problem of cadastral obsolescence in Colombia and its impact on urban planning, tax equity, and territorial management. Within this framework, it explores the capabilities of the foundation model Segment Anything Model v2 (SAM-2) for the automatic segmentation of urban façades using Google Street View imagery, considering façades as key elements in the estimation of real estate market value. The methodological approach included two phases: binary segmentation (façade vs. non-façade) and multiclass segmentation (façade, doors, windows, and meters), assessed through quantitative metrics such as IoU, Dice, precision, recall, and accuracy, complemented with qualitative expert validation. Results show that SAM-2 performs well in general façade identification, although it faces limitations with small or less represented elements in the images. The study concludes that foundation computer vision models provide valuable support for mass appraisal processes within the Multipurpose Cadastre framework, enhancing objectivity, scalability, and cost-efficiency in data collection. Finally, recommendations and future research lines are proposed, aiming to scale these techniques towards material classification, conservation state analysis, and the extraction of geometric metrics relevant for property valuation.