Although 2D architectural floor plans are a commonly used way to express the design of a building, 3D models provide precious insight into modern building usability and safety. In addition, for a historical monument like the Palace of Versailles, the 3D models of its ancient floor plans help us to reconstruct the evolution of its buildings over the years. Such old floor plans are hand made and thus present some problems in automatic creation of a 3D model due to the drawing style variability. In this paper, we introduce a fully automatic and fast method to compute 3D building models from a set of architectural floor plans of the Palace of Versailles dated to the 17 th and 18 th century. First, we detect and localise walls in an input floor plan image using a statistical image segmentation model based on the U-net convolutional neural network architecture and a binary wall mask image is obtained. Secondly, using the generated wall mask image, the 3D model is built upon the linear edge segments representing the detected wall sides in the mask image. In order to cope with the lack of accurate ground truth information for the 3D models of ancient floor plans, we use a dedicated semi-automatic software to build a set of reference 3D models that describe plans' wall projections from three sides of view. We evaluate the performance of our approach on an input floor plan image by measuring the overlapping between the 3D reference model and our 3D model. Our fast and fully automatic approach performs efficiently and produces quite accurate 3D models with 84.2% of IoU score in average. Furthermore, its performance surpasses the performance of the state of the art approach in the wall detection task.