RISC-Net : rotation invariant siamese convolution network for floor plan image retrieval

Abstract

A floor plan represents the blue print of a building. Organizing a massive set of such floor plans and accessing them based on similarity is challenging for any architect. During the digitization process printed floor plan images are rotated slightly by a small degree of angle. Handcrafted feature-based methods proposed in the literature fail to generalize on such scenarios efficiently. In this paper we propose a deep learning-based model, Rotation Invariant Siamese Convolution Network (RISC-Net), which is able to retrieve similar floor plan images from the dataset, even in the presence of rotation. Uniqueness of RISC-Net is the ability to handle scan-time rotation both in the query as well as the images in the database. The proposed method is trained and evaluated on a publicly available floor plan image ROBIN dataset and achieved the best retrieval results 79% as compared to the state-of-the-art methods proposed in the same problem domain.