Kernelized Subspace Ranking for Saliency Detection
In this paper, we propose a novel saliency method that takes advantage of object-level proposals and region-based convolutional neural network (R-CNN) features. We follow the learning-to-rank methodology, and solve a ranking problem satisfying the constra
- PDF / 4,932,732 Bytes
- 17 Pages / 439.37 x 666.142 pts Page_size
- 21 Downloads / 220 Views
Abstract. In this paper, we propose a novel saliency method that takes advantage of object-level proposals and region-based convolutional neural network (R-CNN) features. We follow the learning-to-rank methodology, and solve a ranking problem satisfying the constraint that positive samples have higher scores than negative ones. As the dimensionality of the deep features is high and the amount of training data is low, ranking in the primal space is suboptimal. A new kernelized subspace ranking model is proposed by jointly learning a Rank-SVM classifier and a subspace projection. The projection aims to measure the pairwise distances in a low-dimensional space. For an image, the ranking score of each proposal is assigned by the learnt ranker. The final saliency map is generated by a weighted fusion of the top-ranked candidates. Experimental results show that the proposed algorithm performs favorably against the state-of-the-art methods on four benchmark datasets. Keywords: Saliency detection
1
· Subspace ranking · Feature projection
Introduction
The task of saliency detection is to identify the most attractive and informative regions in images and videos. It has gained much popularity in recent years, owing to its series of important applications in computer vision, such as adaptive compression, context-aware image editing and image resizing. An effective saliency model can save lots of unnecessary human labour in vision tasks. Although much progress in saliency detection has been made in recent years, it remains a challenging problem. The early works [19,21] exploit the low-level image properties of pixels, such as intensity, color, orientation, texture and motion, to compute saliency. Numerous region-wise saliency methods [10,13,43] are proposed subsequently, which investigate the mid-level structure properties of image regions and incorporate the contextual information to measure the saliency for each region. The aforementioned works, either in the pixel-wise or region-wise fashion, have to fully consider the relationship between image elements from overall and local perspectives to guarantee the semantic completeness of salient objects. In this work, we c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part VIII, LNCS 9912, pp. 450–466, 2016. DOI: 10.1007/978-3-319-46484-8 27
Kernelized Subspace Ranking for Saliency Detection
451
explore the category-independent object characteristics of region proposals, and propose a principled framework to weight and combine these region candidates, thereby highlighting the salient instances. Object proposals technique has been widely applied to many vision fields. It generally produces either bounding box proposals [2,8] which inevitably aggregate visual information from objects and background clutter, or region proposals [4,11,32] that shape an informative and well-defined contour. This technique, striving to find instances of all categories, usually produces thousands of object candidates which significantly reduce the search space of salient
Data Loading...
