Increasing Revenue Collection with Computer Vision: Experiments in Pakistan

Economic growth in developing countries is often constrained by the state's inability to generate sufficient tax revenue. A major challenge in improving one key revenue source—property tax—is inaccurate property tax assessments, which can lead to tax leakages and unfair, unequal tax burdens. Computer vision has the potential to improve tax assessment by predicting property valuations from visual data. This study seeks to examine how to leverage computer vision for property tax assessment in two steps. First, working with hundreds of thousands of property images from Punjab, Pakistan, we will develop and refine a computer vision machine learning algorithm that can use property images to predict property assessments. Second, we will conduct a large-scale randomized trial in partnership with the tax authority to evaluate how this algorithm can be used to improve the efficiency and equity of property tax collection.