Abstract
This paper presents a new approach to recognize elements in floor planlayouts. Besides walls and rooms, we aim to recognize diverse floor planelements, such as doors, windows and different types of rooms, in the floorlayouts. To this end, we model a hierarchy of floor plan elements and design adeep multi-task neural network with two tasks: one to learn to predictroom-boundary elements, and the other to predict rooms with types. Moreimportantly, we formulate the room-boundary-guided attention mechanism in ourspatial contextual module to carefully take room-boundary features into accountto enhance the room-type predictions. Furthermore, we design across-and-within-task weighted loss to balance the multi-label tasks andprepare two new datasets for floor plan recognition. Experimental resultsdemonstrate the superiority and effectiveness of our network over thestate-of-the-art methods.