Computer Vision News - November 2016

4 Computer Vision News Research Research Convolutional Oriented Boundaries (COB) Every month, Computer Vision News reviews a research from our field. This month we have chosen to review Convolutional Oriented Boundaries (COB) , a research paper providing a fast and accurate Hierarchical Image Segmentation algorithm, that results in state-of-the-art boundaries, regions and object proposals. The paper was presented only 3 weeks ago at ECCV 2016 , where we had the chance to meet some of the authors ( Kevis-Kokitsi Maninis , Jordi Pont- Tuset , Pablo Arbeláez and Luc Van Gool), who were so kind to concede us an interview which you can read at page 8. The full paper (including code and datasets) is here . Background: Recent advances in Convolutional Neural Networks (CNNs) have produced a dramatic change in performance in many fields of computer vision. Specifically, in the field of contour detection, it has led to the recent appearance of highly accurate systems that rely on large-scale data-sets. The authors of this paper present Convolutional Oriented Boundaries (COB), which produces multiscale oriented contours and region hierarchies. COB has proven to be computationally efficient, since it requires only a single CNN forward pass for contour detection and uses a novel sparse boundary representation for hierarchical segmentation; it provides a significant leap in performance over current state-of-the-art methods, and it generalizes very well to unseen categories and datasets. Extensive experiments performed on BSDS, PASCAL Context, PASCAL Segmentation, and MS-COCO, show that COB provides state- of-the-art contours, region hierarchies, and object proposals. Motivation: For contour detection, the standard method was based on performing several passes of the image in the CNN in different scales. However, it’s well acknowledged that as one moves deeper in the CNN, feature maps capture more global information (this phenomenon also demonstrated by DPM-CNN , which we reviewed in Computer Vision News of October). Thus, the motivation of this work is to take advantage of this intrinsic multiscale nature of the CNN for building an efficient and robust multiscale contour detector obviating the need to perform several passes of the image in the CNN. Novelty: The COB method suggests several novelties and advantages over existing methods: First, it efficiently exploits the multi-scale nature of the CNN architecture by extracting multi-scale features on a single pass. Second, the method estimates not only contour strength but also their orientation. Third,