Citation Context

...proaches can be roughly categorized into two groups: bottom-up methods and topdown methods. Bottom-up methods can be further divided into two categories, i.e. local and global. Local approaches (e.g. =-=[22, 17, 36]-=-) design saliency cues by considering the contrast between each image element (pixel, region, or patch) and its locally surrounding neighborhood. Global approaches estimate saliency scores by calculat...

Citation Context

...tten digit classification and face detection. Recently, the latest generation of CNNs have substantially outperformed handcrafted approaches in computer vision field. Notably, best performing entries =-=[29, 46, 59, 50, 42, 43]-=- on ImageNet ILSVRC [14] and PASCAL VOC [15] benchmarks are all variants of deep CNNs since 2012. Some recent approaches close to our work included cascaded stages in deep learning to solve problems t...

Citation Context

...text modeling, and α, β are parameters of an ambiguity modeling function controlling the need of localcontext modeling. Thus, the posterior probability in Eq.(2) is factorized into product of experts =-=[19]-=-, i.e. we aim to infer the label probability via two components simultaneously: P (y = j | xgc, xlc; θj) ∝ Φ(xgc; θΦj ) ·Ψ(xgc, xlc; θΨj ), (3) θΦj = wgc,j , θ Ψ j = { wgc,j , wlc,j , α, β}, j ∈ {0, 1...

Citation Context

.... Recently, the latest generation of CNNs have substantially outperformed handcrafted approaches in computer vision field. Notably, best performing entries [29, 46, 59, 50, 42, 43] on ImageNet ILSVRC =-=[14]-=- and PASCAL VOC [15] benchmarks are all variants of deep CNNs since 2012. Some recent approaches close to our work included cascaded stages in deep learning to solve problems that need meticulous refi...

Citation Context

...st generation of CNNs have substantially outperformed handcrafted approaches in computer vision field. Notably, best performing entries [29, 46, 59, 50, 42, 43] on ImageNet ILSVRC [14] and PASCAL VOC =-=[15]-=- benchmarks are all variants of deep CNNs since 2012. Some recent approaches close to our work included cascaded stages in deep learning to solve problems that need meticulous refinement, such as in f...

Citation Context

...salient in high-level cognition, i.e. they are distinct in object categories. Therefore, saliency detection is considered as a high-level task in our work. The deep Convolutional Neural Network (CNN) =-=[30]-=-, which recently showed its powerfulness in extracting highlevel feature representations [16], can well solve aforementioned problems. From another perspective, saliency detection is a task to simulat...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...ive attentions in recent years. It has a wide range of applications in computer vision and image processing tasks, such as image/video compression and summarization [38], content-aware image resizing =-=[6]-=-, and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches [6...

Citation Context

...ecific pre-training scheme, as introduced in Section 3.3. 4.5. Evaluation on Overall Performance In Table 2, we compare our approach with nine latest state-of-the-art methods, including IS [20], GBVS =-=[17]-=-, SF [44], GC [12], CEOS [40], PCAS [41], GBMR [57], HS [56], and DRFI [25]. Our approach significantly outperforms all the state-of-the-art salient object segmentation algorithms. The PASCAL-S datase...

Citation Context

...ency detection is considered as a high-level task in our work. The deep Convolutional Neural Network (CNN) [30], which recently showed its powerfulness in extracting highlevel feature representations =-=[16]-=-, can well solve aforementioned problems. From another perspective, saliency detection is a task to simulate the mechanism of human attention, which is a neurocognitive reaction controlled by human br...

Citation Context

...proaches can be roughly categorized into two groups: bottom-up methods and topdown methods. Bottom-up methods can be further divided into two categories, i.e. local and global. Local approaches (e.g. =-=[22, 17, 36]-=-) design saliency cues by considering the contrast between each image element (pixel, region, or patch) and its locally surrounding neighborhood. Global approaches estimate saliency scores by calculat...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...er branch (global-context modeling) of our saliency detection pipeline is a deep CNN architecture with global and coarse context. Superpixel segmentation is firstly performed on images using the SLIC =-=[2]-=- method, and the input of global-context CNN is a superpixel-centered large context window including the full image. Regions exceeding image boundaries are padded with mean pixel value of the training...

Citation Context

...lobal features are weak in capturing semantic information. Top-down methods take advantages of high-level category-specific information as prior knowledge, and are usually task-dependent. Judd et al. =-=[28]-=- learned a top-down saliency model object detectors such as faces, humans, animals, and text. Borji et al. [7] combine bottom-up saliency cues with top-down features learned via multiple object detect...

Citation Context

...tch) and its locally surrounding neighborhood. Global approaches estimate saliency scores by calculating the holistic statistics on uniqueness of each image element over the whole image. Cheng et al. =-=[13, 10]-=- used 3D color histograms as regional features to compute global contrast with all image regions. Perazzi et al. [44] applied two measures of contrast that rate the uniqueness and the spatial distribu...

Citation Context

...oundary and background priors [55, 57, 25, 63]. 2.2. Objectness and Object Proposal Objectness was introduced to measure how likely a region contains an object regardless of object categories. Alexie =-=[3, 4]-=- proposed to combine local appearance contrast and boundary characteristics to measure the objectness score of a bounding box. Based on such measures, some object proposal methods [9, 64] further gene...

Citation Context

...tten digit classification and face detection. Recently, the latest generation of CNNs have substantially outperformed handcrafted approaches in computer vision field. Notably, best performing entries =-=[29, 46, 59, 50, 42, 43]-=- on ImageNet ILSVRC [14] and PASCAL VOC [15] benchmarks are all variants of deep CNNs since 2012. Some recent approaches close to our work included cascaded stages in deep learning to solve problems t...

Citation Context

... fundamental problem drawing extensive attentions in recent years. It has a wide range of applications in computer vision and image processing tasks, such as image/video compression and summarization =-=[38]-=-, content-aware image resizing [6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61...

Citation Context

...tten digit classification and face detection. Recently, the latest generation of CNNs have substantially outperformed handcrafted approaches in computer vision field. Notably, best performing entries =-=[29, 46, 59, 50, 42, 43]-=- on ImageNet ILSVRC [14] and PASCAL VOC [15] benchmarks are all variants of deep CNNs since 2012. Some recent approaches close to our work included cascaded stages in deep learning to solve problems t...

Citation Context

...ories. Alexie [3, 4] proposed to combine local appearance contrast and boundary characteristics to measure the objectness score of a bounding box. Based on such measures, some object proposal methods =-=[9, 64]-=- further generated candidate object regions as a preprocessing step for object detection, which can effectively speed up the process comparing to the classical sliding-window detection paradigm. A rec...

Citation Context

...oundary and background priors [55, 57, 25, 63]. 2.2. Objectness and Object Proposal Objectness was introduced to measure how likely a region contains an object regardless of object categories. Alexie =-=[3, 4]-=- proposed to combine local appearance contrast and boundary characteristics to measure the objectness score of a bounding box. Based on such measures, some object proposal methods [9, 64] further gene...

Citation Context

... images sampled from the MSRA Salient Object Database [36]. Although our training data originates from the same dataset, we separate images in ASD dataset from our training set to avoid overlap. SED1 =-=[5]-=- contains 100 images of a single salient object annotated manually by three users. SED2 [5] contains 100 images of two salient objects annotated manually by three users. ECSSD [56] contains 1, 000 str...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...with taskspecific pre-training scheme, as introduced in Section 3.3. 4.5. Evaluation on Overall Performance In Table 2, we compare our approach with nine latest state-of-the-art methods, including IS =-=[20]-=-, GBVS [17], SF [44], GC [12], CEOS [40], PCAS [41], GBMR [57], HS [56], and DRFI [25]. Our approach significantly outperforms all the state-of-the-art salient object segmentation algorithms. The PASC...

Citation Context

... objects, high-level knowledge on object categories becomes important. Suppose that if the deep model can recognize the gray house, then the problems in Figure 1 can be easily solved. As indicated in =-=[49]-=-, pre-training can provide 1 Image GT Local Context Global Context (a) (b) (c) (d) Figure 2. Examples to show importance of global context. From left to right: image, groundtruth saliency mask, our sa...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...on [38], content-aware image resizing [6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification =-=[62, 61]-=-. A large number of approaches [63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21] are proposed to capture different saliency cues. Many conventional saliency de...

Citation Context

...tten digit classification and face detection. Recently, the latest generation of CNNs have substantially outperformed handcrafted approaches in computer vision field. Notably, best performing entries =-=[29, 46, 59, 50, 42, 43]-=- on ImageNet ILSVRC [14] and PASCAL VOC [15] benchmarks are all variants of deep CNNs since 2012. Some recent approaches close to our work included cascaded stages in deep learning to solve problems t...

Citation Context

...nts of deep CNNs since 2012. Some recent approaches close to our work included cascaded stages in deep learning to solve problems that need meticulous refinement, such as in facial landmark detection =-=[48]-=- and human pose estimation [51]. Saliency detection also need such refinement since global-context model cannot well capture the very detailed information in local neighborhoods. However, we propose a...

Citation Context

...n 3.3. 4.5. Evaluation on Overall Performance In Table 2, we compare our approach with nine latest state-of-the-art methods, including IS [20], GBVS [17], SF [44], GC [12], CEOS [40], PCAS [41], GBMR =-=[57]-=-, HS [56], and DRFI [25]. Our approach significantly outperforms all the state-of-the-art salient object segmentation algorithms. The PASCAL-S dataset was proposed in CPMCGBVS [32], but we do not incl...

Citation Context

...me recent approaches close to our work included cascaded stages in deep learning to solve problems that need meticulous refinement, such as in facial landmark detection [48] and human pose estimation =-=[51]-=-. Saliency detection also need such refinement since global-context model cannot well capture the very detailed information in local neighborhoods. However, we propose a multi-context deep model to co...

Citation Context

...ories. Alexie [3, 4] proposed to combine local appearance contrast and boundary characteristics to measure the objectness score of a bounding box. Based on such measures, some object proposal methods =-=[9, 64]-=- further generated candidate object regions as a preprocessing step for object detection, which can effectively speed up the process comparing to the classical sliding-window detection paradigm. A rec...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...t, and has much stronger generalization capability. In addition, some other interesting priors were also proposed to assist saliency detection, such as flash cues [18], boundary and background priors =-=[55, 57, 25, 63]-=-. 2.2. Objectness and Object Proposal Objectness was introduced to measure how likely a region contains an object regardless of object categories. Alexie [3, 4] proposed to combine local appearance co...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...on [38], content-aware image resizing [6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification =-=[62, 61]-=-. A large number of approaches [63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21] are proposed to capture different saliency cues. Many conventional saliency de...

Citation Context

...t years. It has a wide range of applications in computer vision and image processing tasks, such as image/video compression and summarization [38], content-aware image resizing [6], and photo collage =-=[53]-=-. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches [63, 52, 40, 39, 32, 35, 6...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...ory-specific information as prior knowledge, and are usually task-dependent. Judd et al. [28] learned a top-down saliency model object detectors such as faces, humans, animals, and text. Borji et al. =-=[7]-=- combine bottom-up saliency cues with top-down features learned via multiple object detectors. Yang et al. [58] proposed a top-down saliency model by jointly learning a Conditional Random Field and a ...

Citation Context

...wn saliency model object detectors such as faces, humans, animals, and text. Borji et al. [7] combine bottom-up saliency cues with top-down features learned via multiple object detectors. Yang et al. =-=[58]-=- proposed a top-down saliency model by jointly learning a Conditional Random Field and a dictionary. These methods explore high-level information from 3 ∼ 5 object categories. However, our deep models...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...e/video compression and summarization [38], content-aware image resizing [6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection =-=[37]-=-, and person re-identification [62, 61]. A large number of approaches [63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21] are proposed to capture different salie...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...r of parameters in the above model is about 58 million. We refer readers to [59] for further details. Apart from the Clarifai model, there are also other contemporary models such as AlexNet [29], NIN =-=[33]-=-, OverFeat [46], DeepID-Net [42], and GoogLeNet [50]. It is flexible to incorporate any of these contemporary deep models into our framework, and in the experimental section we investigate the perform...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...on 1, 000 object classes from ImageNet, and has much stronger generalization capability. In addition, some other interesting priors were also proposed to assist saliency detection, such as flash cues =-=[18]-=-, boundary and background priors [55, 57, 25, 63]. 2.2. Objectness and Object Proposal Objectness was introduced to measure how likely a region contains an object regardless of object categories. Alex...

Citation Context

...tten digit classification and face detection. Recently, the latest generation of CNNs have substantially outperformed handcrafted approaches in computer vision field. Notably, best performing entries =-=[29, 46, 59, 50, 42, 43]-=- on ImageNet ILSVRC [14] and PASCAL VOC [15] benchmarks are all variants of deep CNNs since 2012. Some recent approaches close to our work included cascaded stages in deep learning to solve problems t...

Citation Context

...generated candidate object regions as a preprocessing step for object detection, which can effectively speed up the process comparing to the classical sliding-window detection paradigm. A recent work =-=[54]-=- proposed to extract generic objects by jointly handling localization and segmentation tasks. Different than object proposal, which enumerates preliminarily likely candidates for object detection rega...

Citation Context

...f using deep/hierachical architectures to model visual saliency. Yan et al. [56] presented a hierarchical framework to reduce the influence of small-scale structures in saliency detection. Lin et al. =-=[34]-=- proposed to unsupervisedly learn a set of mid-level filters to capture local contrast, and to fuse multi-level saliency calculation by convolution. Unlike their methods where mid-level filters are ha...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...

Citation Context

...6], and photo collage [53]. Saliency information has also been exploited in high-level vision tasks, such as object detection [37], and person re-identification [62, 61]. A large number of approaches =-=[63, 52, 40, 39, 32, 35, 60, 57, 56, 47, 41, 31, 27, 25, 24, 23, 11, 44, 17, 8, 13, 1, 21]-=- are proposed to capture different saliency cues. Many conventional saliency detection methods focus on design of low-level saliency cues, or modeling background GT HS PCAS SF Image Ours DRFI GBMR Fig...