OSED: Object-specific edge detection

Abstract

Object-specific edge detection (OSED) aims to detect object edges in an image along with classify the edge into object or non-object. It prunes edges which are not belonging to the object class for following processing, such as, feature matching for object detection, localization and three-dimensional reconstruction. In this paper, an OSED method that combines region proposal detectors with deep supervision nets to identify object-specific edges is proposed. It minimizes errors of object proposal by learning from hidden layers. Additionally, it combines features from different scales to detect object edges. In order to evaluate the performance of the OSED, we present two datasets which are captured in real scenes. The OSED method demonstrates a high accuracy of 90% and a high speed of 0.5 s for an image whose size is 512 × 448 pixels on the proposed datasets.