REGION-BASED FULLY CONVOLUTIONAL NETWORKS FOR VERTICAL CORNER LINE DETECTION

Region-based Fully Convolutional Network (R-FCN) is recently developed to detect objects and has been successfully used to detect various kinds of semantic objects such as humans and dogs. We investigate the ability of R-FCN for detecting unusual objects in this paper. In detail, based on R-FCN we present an efficient method for vertical corner line (VCL) detection on buildings. Traditionally, VCLs are treated as one kind of image features. In this work, however, they are treated as a class of symbolic objects. Experimental results show that R-FCN, when employed to detect VCLs, could perform potentially better than traditional feature detection algorithms.