Similar floor plan retrieval featuring multi-task learning of layout type classification and room presence prediction

Abstract

In this paper, a new framework for real estate property searches is presented in which a floor plan image is used as a query. In similar property searches, appearance-based similar image retrieval does not work well because similar properties have totally different floor plan images. Therefore, a multitask learning method using deep neural networks to solve this problem is presented. Convolutional Neural Networks (CNNs) are trained to solve the two tasks: layout type classification and room presence classification. The feature vectors obtained from the CNNs are then applied to the retrieval task. Experiments using 22,140 floor plan images in Tokyo, Japan were conducted, and the proposed method achieved the best performance (15.7% with precision@5) compared to other possible approaches.