Treffer: Automated Shelter Recognition in Refugee Camps

In June 2018, more than 68.5 Million people across the globe were reported to be fleeing war or persecution. Within the United Nations, UNOSAT is the organ in charge of collecting demo- graphic information based on satellite images of refugee camps to ensure reliable UN operations providing shelter, food and medicines to refugees and internally displaced people. This work aims at assisting UNOSAT analysts by providing them with an automated shelter detector to increase their efficiency and quality of the analysis. The tool is developed in two phases. The first one consists of generating polygon masks of shelters using single point location generated by UNOSAT analysts in the past 10 years. The second uses the newly generated dataset to develop a model detecting multi-class shelters in new camp images without any analyst data. This report describes the model developed for the first phase of the project. It consists of a fine- tuned Mask R-CNN conditioned on analyst point shelter locations. On average, the tool achieves a recall of 88% and a precision of 81% compared to a human annotated shelter. A user-friendly interface allows the usage of the tool with a command line one-liner. Its usage is summarized in the report and detailed in the Github Repository. The latter also contains the entire python code of the developed tool. The results of all experiments are reported and the weaknesses of the model are discussed. Finally, possible further improvements and next steps required to achieve the development of the automated shelter detector tool are reported.