AFM-RNN: A Sequent Prediction Model for Delineating Building Rooftops from Remote Sensing Images by Integrating RNN with Attraction Field Map

Zeping Liu, Hong Tang, Wei Huang

Abstract: Accurately and automatically delineating the rooftops of buildings from remote sensing images is very essential to many fields. Recently, several Sequent Prediction Models have been proposed to deal with this task. These models use CNN encoder to recognize the boundary fragments by edge and vertex probability maps in order to guide RNN decoder find a set of sequent vertexes linking the boundaries into object regions. However, the recognition result of edge looks like large buffers around the real edges or vertexes of an object instance, which significantly influences the accuracy of predicted polygons. In this paper, we present a novel framework named AFM-RNN, in which a dual representation of lines called Attraction Field Map (AFM) is embedded by Neural Discriminative Dimensionality Reduction Layer (NDDR Layer). Consequently, the problem of over-smooth edge recognition can be effectively solved by directly modeling line segments. Experimental results over three datasets show that the proposed method outperforms other closely related methods. Code is available at https://github.com/Zeping-Liu/AFMRNN-pytorch.