Segmentation-Based Salient Object Detection
Salient object detection is an important task for both the human perception and computer vision applications. Contrary to the popular pixel- or superpixel-based salient object detection methods, we employ high quality segmentation to facilitate salient ob
- PDF / 893,175 Bytes
- 9 Pages / 439.37 x 666.14 pts Page_size
- 62 Downloads / 280 Views
Abstract. Salient object detection is an important task for both the human perception and computer vision applications. Contrary to the popular pixel- or superpixel-based salient object detection methods, we employ high quality segmentation to facilitate salient object detection in this paper. After segmenting the input image using a recent method of gPb-owt-ucm, we easily extract the salient objects from candidate segments only with some very simple intrinsic features of the segments. In addition, for more reasonable performance evaluation, we build a perception based dataset, which contains 499 complex natural images and the corresponding hierarchical salient object ground-truth defined with the assistance of eye-tracker recorded fixations. Experiments on the public ASD dataset and our new dataset show that our segmentation-based salient object detection method (SBSO) achieves competitive performance comparing to some state-of-the-art algorithms. Keywords: Segmentation · Salient object · Fixation · Saliency map
1
Introduction
Saliency computation is an important step which accelerates the computer vision system to efficiently understand natural scenes. There are mainly two categories methods involving in saliency-related visual tasks: (1) fixation prediction methods [1-3], which are usually used to cover the most interesting information of visual scenes; (2) salient object detection methods [4, 5], which are developed mainly to detect the dominant objects from the given images. These two saliency-related tasks are meaningful for various computer vision applications such as image compression [6], image retargeting [7], object recognition [8], etc. Fixation prediction is aimed at predicting where people look when observing a natural scene. This selective attention is considered as a key step for coding and compressing visual information for visual perception [9] (see [10] for a review). Methods for fixation prediction try to obtain a saliency map which indicates the regions of interest in the given images, but usually missing the accurate shape information of salient objects or regions. In contrast, salient object detection usually requires to label pixel-accurate object silhouettes and is used to detect dominant objects in simple scenes [5, 11]. Recent years, a lot of salient object detection methods are developed, © Springer-Verlag Berlin Heidelberg 2015 H. Zha et al. (Eds.): CCCV 2015, Part I, CCIS 546, pp. 94–102, 2015. DOI: 10.1007/978-3-662-48558-3_10
Segmentation-Based Salient Object Detection
95
and most of which detect salient objects based on region contrast. They usually segment the image into lots of small regions with over-segmentation or super-pixel methods, and subsequently, local or global contrast is computed to indicate the saliency level of each region [12-14] (see [4] for a recent review). Recent years, great advance is made in the area of image segmentation, and some methods have achieved excellent performance even for complex images [15, 16]. In this paper, we attempt to clarify how segme
Data Loading...
