Citation Context

...e low-level cues, such as contrast and spectral information, to recognize the most salient regions without realizing content or specific prior knowledge about the targets. The representatives include =-=[3, 6, 8, 9, 11, 13, 14, 15, 17, 20]-=-. Recently, propagation methods have gained much popularity in bottom-up saliency detection and achieved state-ofthe-art performance. To conduct saliency propagations, an input image is represented by...

Citation Context

...nformativity of a superpixel si∈C is straightforwardly modelled by the conditional entropy H(si|L), namely: INF i = H(si|L). (5) The propagations on the graph follow the multivariate Gaussian process =-=[31]-=-, with the elements fi (i=1,· · · , N ) in the random vector f = ( f1 · · · fN )T denoting the saliency values of superpixels si. The associated covariance matrix K equals to the adjacency matrix W ex...

Citation Context

...ixels afterwards, so that the difficult regions can be precisely discovered. Such a “starting simple” strategy conforms to the widely acknowledged theoretical results in pedagogic and cognitive areas =-=[4, 12, 22]-=-, which emphasize the importance of teachers for human’s acquisitions of knowledge from the childish stage to the mature stage. By taking advantage of these psychological opinions, we propose a novel ...

Citation Context

...e low-level cues, such as contrast and spectral information, to recognize the most salient regions without realizing content or specific prior knowledge about the targets. The representatives include =-=[3, 6, 8, 9, 11, 13, 14, 15, 17, 20]-=-. Recently, propagation methods have gained much popularity in bottom-up saliency detection and achieved state-ofthe-art performance. To conduct saliency propagations, an input image is represented by...

Citation Context

...e the proposed Teaching-to-Learn and Learning-toTeach approach (abbreviated as “TLLT”) with twelve popular methods on two popular saliency datasets. The twelve baselines include classical methods (LD =-=[16]-=-, GS [23]), state-of-the-art methods (SS [6], PD [17], CT [13], RBD [30], HS [25], SF [20]), and representative propagation based methods (MR [27], GP [5], AM [10], GRD [26]). The parameters in our me...

Citation Context

...Harris corner detector. Because most key points locate within the target region, we link the outer key points to a convex hull to roughly enclose the target (see Fig. 2). We proceed by using the SLIC =-=[1]-=- algorithm to oversegment the input image into N small superpixels (see Fig. 2), then an undirected graph G = 〈V, E〉 is built where V is the node set consisted of these superpixels and E is the edge s...

Citation Context

...e low-level cues, such as contrast and spectral information, to recognize the most salient regions without realizing content or specific prior knowledge about the targets. The representatives include =-=[3, 6, 8, 9, 11, 13, 14, 15, 17, 20]-=-. Recently, propagation methods have gained much popularity in bottom-up saliency detection and achieved state-ofthe-art performance. To conduct saliency propagations, an input image is represented by...

Citation Context

...e low-level cues, such as contrast and spectral information, to recognize the most salient regions without realizing content or specific prior knowledge about the targets. The representatives include =-=[3, 6, 8, 9, 11, 13, 14, 15, 17, 20]-=-. Recently, propagation methods have gained much popularity in bottom-up saliency detection and achieved state-ofthe-art performance. To conduct saliency propagations, an input image is represented by...

Citation Context

...ysical Interpretation and Justification A key factor to the effectiveness of our method is the well-ordered learning sequence from simple to difficult, which is also considered by curriculum learning =-=[1]-=- and self-paced learning [8]. Our paper introduces this strategy to graph-based saliency propagation. More interestingly, we provide a physical interpretation of this strategy, by relating the curricu...

Citation Context

...milar to the labeled ones. For example, [7] and [10] formulate the saliency propagation process as random walks on the graph. [21] and [27] conduct the propagations by employing personalized PageRank =-=[29]-=- and manifold based diffusion [29], respectively. All these methods generate similar propagation sequences which are heavily influenced by the superpixels’ spatial relationships. However, once encount...

Citation Context

...ixels afterwards, so that the difficult regions can be precisely discovered. Such a “starting simple” strategy conforms to the widely acknowledged theoretical results in pedagogic and cognitive areas =-=[4, 12, 22]-=-, which emphasize the importance of teachers for human’s acquisitions of knowledge from the childish stage to the mature stage. By taking advantage of these psychological opinions, we propose a novel ...

Citation Context

...lements are set to 1. 2For simplicity, the superscript t is omitted for all the notations hereinafter unless otherwise specified. For the multivariate Gaussian, the closed-form solution of H(si|L) is =-=[2]-=-: H(si|L) = 1 2 ln(2pieσ2i|L), (6) where σ2i|L denotes the conditional covariance of fi given L. Considering that the conditional distribution is a multivariate Gaussian, σ2i|L in (6) can be represent...

Citation Context

...e low-level cues, such as contrast and spectral information, to recognize the most salient regions without realizing content or specific prior knowledge about the targets. The representatives include =-=[3, 6, 8, 9, 11, 13, 14, 15, 17, 20]-=-. Recently, propagation methods have gained much popularity in bottom-up saliency detection and achieved state-ofthe-art performance. To conduct saliency propagations, an input image is represented by...

Citation Context

...Up to now, a great number of detectors based on computaFigure 1. The results achieved by typical propagation methods and our method on two example images. From left to right: input images, results of =-=[27]-=-, [10], and our method. tional intelligence have been proposed. They can be roughly divided into two categories: bottom-up methods that are data and stimulus driven, and top-down methods that are task...

Citation Context

...LLT”) with twelve popular methods on two popular saliency datasets. The twelve baselines include classical methods (LD [16], GS [23]), state-of-the-art methods (SS [6], PD [17], CT [13], RBD [30], HS =-=[25]-=-, SF [20]), and representative propagation based methods (MR [27], GP [5], AM [10], GRD [26]). The parameters in our method are set to N = 400 and θ= 0.25 throughout the experiments. The parametric se...

Citation Context

...the background propagations. The first one is the convex hull prior [26] that assumes the pixels outside the convex hull are very likely to be the background; and the second one is the boundary prior =-=[23, 27]-=- which indicates the regions along the image’s four boundaries are usually non-salient. Note that the two priors are not strong because the specified non-salient regions will be further refined or 1In...

Citation Context

...agation. More interestingly, we provide a physical interpretation of this strategy, by relating the curriculum guided propagation to the practical fluid diffusion. In physics, Fick’s Law of Diffusion =-=[2]-=- is well-known for understanding the mass transfer of solids, liquids, and gases through diffusive means. It postulates that the flux diffuses from regions of high concentration to regions of low conc...

Citation Context

...e low-level cues, such as contrast and spectral information, to recognize the most salient regions without realizing content or specific prior knowledge about the targets. The representatives include =-=[3, 6, 8, 9, 11, 13, 14, 15, 17, 20]-=-. Recently, propagation methods have gained much popularity in bottom-up saliency detection and achieved state-ofthe-art performance. To conduct saliency propagations, an input image is represented by...

Citation Context

...ixels afterwards, so that the difficult regions can be precisely discovered. Such a “starting simple” strategy conforms to the widely acknowledged theoretical results in pedagogic and cognitive areas =-=[4, 12, 22]-=-, which emphasize the importance of teachers for human’s acquisitions of knowledge from the childish stage to the mature stage. By taking advantage of these psychological opinions, we propose a novel ...

Citation Context

...s the non-target regions, so the real targets are not precisely detected in the final saliency map (Fig. 10(d)). However, this case seldom occurs according to the extensive experiments in prior works =-=[5, 24, 26]-=-. The failure rates of the convex (a) (b) (c) (d) Figure 10. Failed cases of our method. (a) shows an example that the object is very similar to the background, in which the correct seed superpixels a...

Citation Context

...now, a great number of detectors based on computaFigure 1. The results achieved by typical propagation methods and our method on two example images. From left to right: input images, results of [27], =-=[10]-=-, and our method. tional intelligence have been proposed. They can be roughly divided into two categories: bottom-up methods that are data and stimulus driven, and top-down methods that are task and k...

Citation Context

...e low-level cues, such as contrast and spectral information, to recognize the most salient regions without realizing content or specific prior knowledge about the targets. The representatives include =-=[3, 6, 8, 9, 11, 13, 14, 15, 17, 20]-=-. Recently, propagation methods have gained much popularity in bottom-up saliency detection and achieved state-ofthe-art performance. To conduct saliency propagations, an input image is represented by...

Citation Context

...ed superpixels to their unlabeled neighbors. However, such 1 propagations may incur errors if the unlabeled adjacent superpixels are inhomogeneous or very dissimilar to the labeled ones. For example, =-=[7]-=- and [10] formulate the saliency propagation process as random walks on the graph. [21] and [27] conduct the propagations by employing personalized PageRank [29] and manifold based diffusion [29], res...

Citation Context

...ted as “TLLT”) with twelve popular methods on two popular saliency datasets. The twelve baselines include classical methods (LD [16], GS [23]), state-of-the-art methods (SS [6], PD [17], CT [13], RBD =-=[30]-=-, HS [25], SF [20]), and representative propagation based methods (MR [27], GP [5], AM [10], GRD [26]). The parameters in our method are set to N = 400 and θ= 0.25 throughout the experiments. The para...

Citation Context

...e low-level cues, such as contrast and spectral information, to recognize the most salient regions without realizing content or specific prior knowledge about the targets. The representatives include =-=[3, 6, 8, 9, 11, 13, 14, 15, 17, 20]-=-. Recently, propagation methods have gained much popularity in bottom-up saliency detection and achieved state-ofthe-art performance. To conduct saliency propagations, an input image is represented by...

Citation Context

...meters in our method are set to N = 400 and θ= 0.25 throughout the experiments. The parametric sensitivity and failed cases are also discussed at the end of this section. 4.1. Metrics Margolin et al. =-=[18]-=- point out that the traditional Precision-Recall curve (PR curve) and Fβ-measure suffer the interpolation flaw, dependency flaw and equalimportance flaw. Instead, they propose the weighted precision P...

Citation Context

...rors if the unlabeled adjacent superpixels are inhomogeneous or very dissimilar to the labeled ones. For example, [7] and [10] formulate the saliency propagation process as random walks on the graph. =-=[21]-=- and [27] conduct the propagations by employing personalized PageRank [29] and manifold based diffusion [29], respectively. All these methods generate similar propagation sequences which are heavily i...

Citation Context

...bottom-up methods that are data and stimulus driven, and top-down methods that are task and knowledge driven. Top-down methods are usually related to the subsequent applications. For example, Maybank =-=[19]-=- proposed a probabilistic definition of salient image regions for image matching. Yang et al. [28] combined dictionary learning and Conditional Random Fields (CRFs) to generate discriminative represen...

Citation Context

...undaries as seeds, and implement the propagation again. A saliency map based on the boundary prior can then be generated, which is denoted as SBoundary . Furthermore, we establish a binary mask Smask =-=[5]-=- to indicate whether the i-th superpixel is inside (SMask(i) = 1) or outside (SMask(i) = 0) the convex hull H. Finally, the saliency map of Stage 1 is obtained by integrating SConvexHull, SBoundary , ...

Citation Context

...e low-level cues, such as contrast and spectral information, to recognize the most salient regions without realizing content or specific prior knowledge about the targets. The representatives include =-=[3, 6, 8, 9, 11, 13, 14, 15, 17, 20]-=-. Recently, propagation methods have gained much popularity in bottom-up saliency detection and achieved state-ofthe-art performance. To conduct saliency propagations, an input image is represented by...

Citation Context

... (see the magenta arrows in Fig. 2), which will be concretely introduced in Section 3. 2.1. Image Pre-processing Given an input image, a convex hull H is constructed to estimate the target’s location =-=[26]-=-. This is done by detecting some key points in the image via Harris corner detector. Because most key points locate within the target region, we link the outer key points to a convex hull to roughly e...