Convolutional Oriented Boundaries
We present Convolutional Oriented Boundaries (COB), which produces multiscale oriented contours and region hierarchies starting from generic image classification Convolutional Neural Networks (CNNs). COB is computationally efficient, because it requires a
- PDF / 3,306,938 Bytes
- 17 Pages / 439.37 x 666.142 pts Page_size
- 4 Downloads / 192 Views
2
ETH Z¨ urich, Z¨ urich, Switzerland [email protected] Universidad de Los Andes, Bogot´ a, Colombia 3 KU Leuven, Leuven, Belgium
Abstract. We present Convolutional Oriented Boundaries (COB), which produces multiscale oriented contours and region hierarchies starting from generic image classification Convolutional Neural Networks (CNNs). COB is computationally efficient, because it requires a single CNN forward pass for contour detection and it uses a novel sparse boundary representation for hierarchical segmentation; it gives a significant leap in performance over the state-of-the-art, and it generalizes very well to unseen categories and datasets. Particularly, we show that learning to estimate not only contour strength but also orientation provides more accurate results. We perform extensive experiments on BSDS, PASCAL Context, PASCAL Segmentation, and MS-COCO, showing that COB provides state-of-the-art contours, region hierarchies, and object proposals in all datasets. Keywords: Contour detection Hierarchical image segmentation
1
· Contour orientation · Object proposals
estimation
·
Introduction
The adoption of Convolutional Neural Networks (CNNs) has caused a profound change and a large leap forward in performance throughout the majority of fields in computer vision. In the case of a traditionally category-agnostic field such as contour detection, it has recently fostered the appearance of systems [1–6] that rely on large-scale category-specific information in the form of deep architectures pre-trained on Imagenet for image classification [7–10]. This paper proposes Convolutional Oriented Boundaries (COB), a generic CNN architecture that allows end-to-end learning of multiscale oriented contours, and we show how it translates top performing base CNN networks into high-quality contours; allowing to bring future improvements in base CNN architectures into semantic grouping. We then propose a sparse boundary representation for efficient construction of hierarchical regions from the contour signal. Our overall approach is both efficient (it runs in 0.8 seconds per image) and highly accurate (it produces state-of-the-art contours and regions on PASCAL and on the BSDS). Figure 1 shows an overview of our system. c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part I, LNCS 9905, pp. 580–596, 2016. DOI: 10.1007/978-3-319-46448-0 35
Convolutional Oriented Boundaries
581
Fig. 1. Overview of COB: From a single pass of a base CNN, we obtain multiscale oriented contours. We combine them to build Ultrametric Contour Maps (UCMs) at different scales and fuse them into a single hierarchical segmentation structure.
For the last fifteen years, the Berkeley Segmentation Dataset and Benchmark (BSDS) [11] has been the experimental testbed of choice for the study of boundary detection and image segmentation. However, the current large-capacity and very accurate models have underlined the limitations of the BSDS as the primary benchmark for grouping. Its 300 train images are inadequate for train
Data Loading...
