Convolutional Network for Generic Object Contour Detection with Stereo Vision

Abstract

Object contour detection is the task of extracting the shape created by the boundaries between objects in an image. The conventional method previously limits the detection targets to specific categories, or detects edges of patterns inside an object. We propose a new method to represent a contour image with a distance to the boundary as a pixel value. We also propose a deep convolutional network for generic object contour detection combined with stereo vision. In the network learning, false positives inside objects are penalized with increasing distance to boundaries. As a result of experiments, the proposed network drew a good precision-recall curve, and F-measure was about 0.7 for Driving dataset. False detections inside the object were significantly reduced. Furthermore, using common network structure and similar inference targets, we propose a network that simultaneously estimates disparity and contour with fully shared weights. This network reduced the amount of calculation and memory capacity by half, and the drop in accuracy was slight.