Uncertainty Guided Polygon Generation for Building Detection

Abstract

Enhanced accuracy of building detection algorithms has the potential to benefit a wide array of applications, including urban planning, environmental monitoring, and disaster response efforts. However, building extraction algorithms struggle with robustness due to among others, occlusions from vegetation and shadows of nearby tall buildings, complex building shapes, and a large distributional shift between datasets that come in varying spatial resolutions, resulting in their dependency of dataset-specific and user specified parameters. Towards addressing this shortcoming, we hypothesize that the model's uncertainty can be leveraged to increase the robustness and efficacy of these algorithms. As a first step towards evaluating this hypothesis, we propose an improved version of the current state-of-the-art that incorporates quantification of the model uncertainty. We further show that leveraging these uncertainty methods by guiding the vertex selection process through the use of a dynamic threshold improves the stability across datasets. Results on two datasets demonstrate that incorporating uncertainty has the potential to significantly improve the robustness of the previous state-of-the-art method. Additionally, the dynamic threshold, while offering a more modest improvement, showcases the potential of actively leveraging uncertainty measures to improve the model performance.