SketchPLAN Recognition and Vectorization of Floor Plan Sketches for Building Information Modelling Design Environment

Our work builds upon general-purpose sketch recognition research, proposing a framework to integrate sketches within architectural design software. The use case being freehand-drawn floor plans as input for the Building Information Modelling environment. For that, we developed the SketchPLAN process, which recognizes, vectorizes, and contextualizes floor plan sketches. For recognition, SketchPLAN leverages the power of conditional Generative Adversarial Networks (pix2pix) to generate semantic segmentation maps out of floor plan sketch inputs, after training it with our own annotated dataset. The next step consists of vectorizing the segmented raster output through a floor plan vectorization library. A graphical user interface was designed to streamline the process. Bringing its output into Revit environment goes through the pipeline of “Rhino.Inside”, converting geometry objects into Revit type instances. SketchPLAN demonstrates how artificial intelligence tools can be harnessed to bring floor plan sketches as input into BIM environments. In terms of contribution, we have collected and annotated an image dataset of floor plan sketches. The annotation system was thought right from the start as a means to vectorize those images. Moreover, our developed floor plan vectorization library performed well in converting those segmentation maps into usable geometry. The proposed solution shows nevertheless some deficiencies: struggling to omit all background noise when recognizing, and not considering some special cases when vectorizing. SketchPLAN holds however a bigger potential, considering the possible improvements on the process in terms of data collection, model architecture, more comprehensive vectorization, as well as implementation in other scenarios such as design recommendation tools.