Automatic text placement on maps using deep learning keypoint detection models

Abstract: Labeling the map is one of the most essential parts of the cartographic process that requires a huge time and energy. It is proven that the automation of map labeling is an NP-hard problem. There have been many research studies that tried to solve it such as rule-based methods, metaheuristics, and integer programming. However, the results achieved so far are not satisfactory and require much manual processing. In fact, many cartographic rules are hard to quantify or formulate as objective function or to include as a constraint. The purpose of this master thesis was to find a new way for text placement and introduce a method based on keypoint detection using deep learning. For this goal, a workflow is designed and consists of rasterization of the manually labeled data, followed by data augmentation and shaping. Then, based on the experiments, the architecture and the parameters of the Stacked Hourglass Networks are determined based on the evaluated performance. The best-found architectures achieved an accuracy of 60.7%. Furthermore, with an attention mechanism, the model can achieve an accuracy of 63.1%.