Over the last years, Open Street Map (OSM) data has been produced for a high percent of the Earth surface most of these data covering developed regions. Even with the increasingly number of volunteers and initiatives like Humanitarian OpenStreetMap (HOT) and Missing Maps, it is hard to believe that the rest of the planet can be mapped in the next years, not mentioning the further requirements for updating the maps, once created or data accuracy that is prone to various errors. As manual digitizing is a complex and questioned process, it might be useful to support it by automated remote sensing approaches. While automated approaches are not expected to deliver fully accurate results, they still could save important work, as correcting errors by editing (like delete false positive buildings) is expected to work faster than digitizing on the scratch. Because, after a deep literature review, there is no building extraction method suitable for our problem, the aim of this study is to develop a buildings extraction algorithm by combining an object-based (OBIA) feature extraction approach with a deep learning algorithm that gives the best results in our days: convolutional neural network (CNN). The results (building polygons) from the OBIA algorithm are used as training samples for CNN algorithm. The algorithm has been tested on RGB images from Sztaki-Inria and Ali Özgün OK building detection benchmarks. In all cases, the algorithm has provided promising results, with an overall accuracy over 85% and very good geometric accuracies. The buildings extraction algorithm will be further tested for transferability and usefulness in supporting the OSM initiative, particularly in areas exposed to humanitarian crisis where OSM data does not exist at all.