GETTING A HANDLE ON FLOOR PLAN ANALYSIS - DOOR CLASSIFICATION IN FLOOR PLANS AND A SURVEY ON EXISTING DATASETS

Abstract Floor plan interpretation and reconstruction is crucial to enable the transformation of drawings to 3D models or different digital formats. It has recently taken advantage of neural-based architectures, especially in the semantic segmentation field. These techniques perform better than traditional methods, but the results depend mainly on the data used to train the networks, which is often crafted for the specific task being performed, making it hard to reuse for different purposes. In this paper, we conduct a literature survey on the existing datasets for floor plan analysis, and we explore how information regarding door placement and orientation can be recovered without having to change the initial data or model. We propose a two-step recognition method based on image segmentation followed by classification of cropped zones to allow data augmentation during training. In the process, we generate a dataset consisting of 35000 annotated door images extracted from an existing dataset.