Contour Loss: Boundary-Aware Learning for Salient Object Segmentation

Huajun Zhou; Jianhuang Lai; Xiaohua Xie; Zixuan Chen

REVIEW 3 major objections 6 minor 1 cited by

Contour Loss: Boundary-Aware Learning for Salient Object Segmentation

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that Contour Loss and a hierarchical global attention module, combined in an FPN-style VGG-16 network, outperform 13 prior methods on six salient-object-segmentation benchmarks.

desk verdict A modest, plausible saliency-detection paper whose Contour Loss claim is real but not directly proven; deserves a referee, not a desk reject. read the letter →

arxiv 1908.01975 v1 pith:XXVXR5GX submitted 2019-08-06 cs.CV

Zixuan Chen , Huajun Zhou , Xiaohua Xie , Jianhuang Lai This is my paper · ORCID

classification cs.CV

keywords salientobjectsegmentationContourLossboundary-awarelearninghierarchicalglobalattentionmodulecontextfeaturepyramidnetworkdeepsupervisionVGG-16

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that salient object segmentation improves when the training signal itself points at object boundaries and when the decoder is steered by global context at every scale. To do that, the authors introduce Contour Loss, a reweighting of binary cross-entropy that gives extra weight to pixels near edges via a dilated-minus-eroded ground-truth map, and the hierarchical global attention module (HGAM), which propagates unsupervised contrast-based attention maps down the feature pyramid. These are combined in an FPN-style network with a VGG-16 backbone. If the claim holds, the result is sharper boundaries and cleaner background suppression than 13 earlier methods across six benchmark datasets, at a speed of 26 frames per second.

What carries the argument

The load-bearing objects are two. The first is Contour Loss: a spatial weight map $M_C=\mathrm{Gauss}(K\cdot((Y;S)_+-(Y;S)_-))+1$, with $K=5$, built from the difference between dilated and eroded ground-truth masks, multiplies the per-pixel binary cross-entropy term so boundary pixels dominate training. The second is HGAM: each module concatenates pooled, encoded, and previous-message features, forms an attention map $H^{\mathrm{Atten}}=\delta((F_{\mathrm{In}}-\operatorname{Aver}(F_{\mathrm{In}}))/\sqrt{\operatorname{Var}(F_{\mathrm{In}})+\epsilon})+\lambda_1$, and multiplies that map onto residual decoder features, passing messages coarse-to-fine so each level sees global contrast rather than a softmax-sparse set of pixels.

What would settle it

A decisive test: retrain the full network twice on DUTS-TR, once with every Contour Loss weight fixed to 1 (plain cross-entropy) and once with HGAM connections severed; if the full model does not beat both ablated variants by roughly 0.01 to 0.02 in max $F_\beta$ on DUTS-TE and DUT-O, the paper's mechanism is not the source of the claimed improvement.

Watch

Extended reading notes

Core claim

On the authors' own terms, the central discovery is that boundary-focused supervision and global contextual attention are complementary and jointly push saliency segmentation past earlier methods. Contour Loss alone raises the baseline's maximum $F_\beta$ on DUTS-TE from 0.848 to 0.861, HGAM alone raises it to 0.860, and the combination reaches 0.872, with analogous MAE gains and consistent improvements on DUT-O. The paper argues that these improvements come from the two modules addressing different error types: Contour Loss makes predicted boundaries match the mask edges, while HGAM removes background distractions that softmax-based attention misses. The reported system also keeps inference at 26 fps, so the accuracy gain is not bought with extra test-time computation.

Load-bearing premise

The boundary-aware gain rests on the assumption that the fixed weight map built from dilation, erosion, and Gaussian smoothing with $K=5$ correctly identifies the pixels whose misclassification matters most; if that proxy does not generalize, the reported improvements would shrink or vanish outside the six datasets.

Editorial extensions

If this is right

Any binary segmentation task can reuse Contour Loss as a drop-in loss term, since it needs only a mask and morphological operations.
HGAM gives high-resolution dense predictors an attention mechanism that retains background information instead of collapsing onto a few pixels.
The 26 fps inference means the accuracy gains come without slowing deployment, making the approach usable in real-time systems.
The ablations imply the two modules attack different error modes, so combining boundary weighting with global-context attention is a reusable recipe rather than a dataset-specific fix.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The authors leave implicit that Contour Loss is not saliency-specific; the same dilation-minus-erosion weight map could be applied to semantic segmentation or medical imaging, where boundary errors dominate.
A testable extension is to swap HGAM into other FPN-style decoders in place of softmax attention; the paper's mechanism predicts the benefit should grow with image resolution.
Because HGAM's attention maps are unsupervised and sometimes emphasize background, they may act as an implicit regularizer; that predicts HGAM helps most on smaller or noisier training sets.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a modest empirical paper, not a breakthrough. The new pieces are a boundary-weighted cross-entropy loss (Contour Loss) and a global-contrast attention module (HGAM), both riding on an FPN/VGG-16 baseline. On six standard saliency benchmarks the full model tops prior VGG-16 methods by 1-2 points Fβ, and runs at 26 fps. I believe the results are probably real, but the paper doesn't prove the boundary mechanism it claims.

What's good: the ablations in Table 2 show each component helps, on two large datasets, and the gains are consistent. The visual examples, while cherry-picked, do show finer contours with Contour Loss. The design is simple enough to reproduce, and the loss could transfer to other dense-prediction tasks. The writing is clear, and the comparisons are fair on the whole.

Where I'd push back: (1) No boundary-specific metric. The paper's title and central claim are about boundaries, but all numbers are global Fβ and MAE. Those can improve from non-boundary effects like better interior filling or a slight shift of the predicted region. The authors never report boundary F-measure, trimap F, or contour precision/recall, so the mechanism is not actually confirmed. A single boundary metric would have settled it. (2) No error bars, no repeated runs, no released code. That's common in this subfield, but the gains are small enough (1-2%) that I'd like to know they're stable. The ablation is also only on DUTS-TE and DUT-O. (3) The attention map in Eq. 6 is under-specified: δ was earlier defined as conv+ReLU, but the attention map is used for multiplicative gating; if it's not bounded to [0,1] that's a problem. A typo ('Guass') and the lack of a weight for the final P loss in Eq. 9 are minor.

Overall: the stress-test note is right but a bit severe. The evidence is suggestive, not conclusive. This isn't a desk-reject paper; it's a reasonable conference submission that would benefit from a boundary-aware evaluation and some robustness checks. I'd send it to review and ask for that.

Referee Report

3 major / 6 minor

Summary. The paper proposes two components for salient object segmentation: Contour Loss, a reweighted cross-entropy loss that up-weights pixels near salient-object boundaries using a spatial weight map derived from dilated/eroded ground-truth masks, and a hierarchical global attention module (HGAM) that captures global context at multiple resolutions. The architecture is an FPN-style VGG-16 network with residual decoder blocks; the final output is generated from the guided first-level feature. The method is evaluated on six benchmarks against 13 prior methods using maximum F-beta and MAE, and an ablation on DUTS-TE and DUT-O compares the baseline with Contour Loss alone, HGAM alone, and their combination. The authors report consistent improvements and a real-time inference speed of 26 fps.

Significance. If the claims hold, the paper makes a modest but useful contribution: a simple boundary-weighting loss and a global-context attention module that together improve salient object segmentation over VGG-16-based competitors. The strengths are the clear formulation of the two components, the breadth of the benchmark comparison, the inclusion of PR curves and visualizations, and the explicit ablation study. The central boundary-aware claim, however, is not directly tested: only whole-image F-beta and MAE are reported, and the gains over the closest baselines are small and shown without repeated runs or significance testing. The reader's concern that the Contour Loss benefit might come from non-boundary effects therefore lands, and it is the main load-bearing issue for the paper's framing.

major comments (3)

[§3.2, §4.3, Table 2] The central claim that Contour Loss makes the network 'boundary-aware' is not directly tested. Table 2 reports only whole-image maximum F-beta and MAE, and Table 1 likewise uses only these global metrics. These metrics can improve from better interior labeling or from a slight shift in the predicted object extent, rather than from more accurate contour localization. Since the loss is named and motivated by boundary perception, the paper should report a boundary-specific metric (for example boundary F-measure, trimap F-measure, or contour precision/recall) on at least DUTS-TE and DUT-O. Without such a metric, the comparison B+C versus B in the ablation does not establish that the improvement is due to improved boundary awareness.
[§4.3, Table 2] The reported improvements are small and are presented without error bars or repeated runs. For instance, on DUTS-TE the maximum F-beta values are 0.848 for B, 0.861 for B+C, 0.860 for B+H, and 0.872 for the full model; the corresponding DUT-O values are 0.787, 0.806, 0.801, and 0.825. Differences of 0.001 to 0.005 between B+C and B+H are within typical run-to-run variation for a single training run. The authors should provide mean and standard deviation over multiple runs, or a statistical significance test, to justify the claim that each component contributes independently and that the combined gain is reliable.
[§3.2, Eq. (4), §4.1] The Contour Loss depends on the empirically chosen K=5 and a fixed 5x5 dilation/erosion mask, while HGAM depends on lambda=0.1 in Eq. (6). No sensitivity analysis is given for these hyperparameters. Because the loss's effect is central to the paper and the reported gains are small, the authors should show that the conclusions are stable over a reasonable range of K, or describe a validation-based selection procedure for K and lambda. Without this, the transferability of the boundary-weighting scheme to other datasets or settings remains unclear.

minor comments (6)

[§3.2, Eq. (4)] The text spells 'Gaussian' as 'Guass' in Eq. (4); also define the structuring element S and the border handling used in the dilation and erosion operations.
[Fig. 7 caption] The caption lists 'DRFI[5]' but the reference list and Table 1 use DRFI[9]; the citation should be corrected.
[§4.4] The heading 'HGAM Visualizaiton' contains a typo and should read 'Visualization'.
[Fig. 2 caption] The caption refers to 'guidedRes1' before the notation ResG_i is defined in Eq. (8); please unify the notation for clarity.
[Table 1] The compared methods use different backbones (VGG-16, VGG-19, ResNet-50); the 'state-of-the-art' claim should be qualified by noting which comparisons are most direct, and the reader should be pointed to the closest VGG-16 competitors.
[§4.1] The paper does not mention code or model release; a reproducibility statement or a promise of releasing the code would strengthen the contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the loss and attention module are defined from ground-truth masks and input features, with all ablations reported on held-out test sets.

full rationale

The paper's derivation chain is self-contained and non-circular. Contour Loss (Eqs. 4-5) is a training objective: a spatial weight map M_C is computed from dilated and eroded ground-truth masks and used to reweight cross-entropy terms. This is a legitimate loss design, not a prediction derived from the quantity it claims to predict. HGAM (Eqs. 6-8) generates attention maps from input feature statistics and pooling operations, then guides residual features; it is trained jointly through the final saliency loss rather than being fitted to the reported test metrics. The final model is evaluated on six benchmarks, and the ablation in Table 2 compares settings such as B, B+C, B+H, and B+C+H on held-out test sets. Hyperparameters such as K=5, lambda=0.1, and the layer weights are described as empirically chosen, and the paper does not state that test-set metrics were used to select them; even if hand-tuned, they are not 'predictions' of those metrics. References to FPN, VGG-16, focal loss, and attention modules are external prior work and are not self-citations carrying the argument. The absence of a boundary-specific evaluation metric is a legitimate evidence-strength concern, but it is an empirical-validation issue, not a circular-derivation issue. No equation reduces to its own input, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical and rests on the chosen hyperparameters and the modeling assumptions above. No new objects, forces, or dimensions are introduced.

free parameters (3)

K = 5
Hyperparameter in Eq. 4 scaling the boundary weight map; set 'empirically' in Section 3.2.
lambda = 0.1
Regularization term in Eq. 6 for HGAM attention; set 'empirically' in Section 3.3.
Hierarchical loss weights W_L = 0.3, 0.4, 0.6, 0.8, 1.0
Weights for intermediate predictions in Eq. 3 and Eq. 9, chosen to reflect coarse-to-fine supervision (Section 4.1).

assumptions (4)

domain assumption Pixels near object boundaries are harder to classify and deserve higher training weight
Motivates Contour Loss in Section 3.2; not derived from data or theory.
domain assumption Global contrast, implemented as mean subtraction and standardization, better captures global saliency than softmax attention
Assumption underlying HGAM in Section 3.3; supported only by the reported experiments.
domain assumption ImageNet-pretrained VGG-16 features transfer to saliency segmentation
The encoder is initialized with ImageNet weights and fine-tuned, as stated in Section 3.1.
domain assumption The reported numerical results for prior methods were obtained under protocols comparable to the authors' evaluation
Table 1 compares against numbers from other papers; any protocol differences would undermine the SOTA claim.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Contour Loss: Boundary-Aware Learning for Salient Object Segmentation." pith.science (2026). https://pith.science/paper/XXVXR5GX

@misc{pith2026190801975,
  author       = {Pith},
  title        = {Pith review of: Contour Loss: Boundary-Aware Learning for Salient Object Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XXVXR5GX}},
  note         = {Machine review of arXiv:1908.01975}
}

read the original abstract

We present a learning model that makes full use of boundary information for salient object segmentation. Specifically, we come up with a novel loss function, i.e., Contour Loss, which leverages object contours to guide models to perceive salient object boundaries. Such a boundary-aware network can learn boundary-wise distinctions between salient objects and background, hence effectively facilitating the saliency detection. Yet the Contour Loss emphasizes on the local saliency. We further propose the hierarchical global attention module (HGAM), which forces the model hierarchically attend to global contexts, thus captures the global visual saliency. Comprehensive experiments on six benchmark datasets show that our method achieves superior performance over state-of-the-art ones. Moreover, our model has a real-time speed of 26 fps on a TITAN X GPU.

Figures

Figures reproduced from arXiv: 1908.01975 by the authors.

**Figure 2.** Overall architecture of the proposed network with VGG-16 [ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Inner structure of the ith HGAM. Yellow, blue and gray dotted line represent attention stream, prediction stream and attention guidance respectively. wi and hi severally denote width and height of Ei . HOut i and HAtten i are the ith HGAM message and attention map respectively. resent the average and variance value of F In respectively. λ denotes a regularization term which is set to 0.1 empirically, while is a… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: PR curves of ours and other state-of-the-art methods. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 6.** Figure 6: Visualization of attention maps [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: Visual comparison between ours and other state-of-the-art methods. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

A Novel Shape Guided Transformer Network for Instance Segmentation in Remote Sensing Images
cs.CV 2024-12 conditional novelty 4.0 of 10

A shape-guided transformer with row and column global attention reaches state-of-the-art instance segmentation accuracy on three remote sensing datasets.

Reference graph

Works this paper leans on

39 extracted references · 38 canonical work pages · cited by 1 Pith paper

[1]

Achanta, A

R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. S ¨usstrunk. Slic superpixels compared to state-of-the- art superpixel methods. IEEE TPAMI, 34(11):2274–2282,

work page
[2]

S. Chen, X. Tan, B. Wang, and X. Hu. Reverse attention for salient object detection. In ECCV, 2018. 3, 6, 7, 9

work page 2018
[3]

Cheng, N

M.-M. Cheng, N. J. Mitra, X. Huang, P. H. Torr, and S.-M. Hu. Global contrast based salient region detection. IEEE TPAMI, 37(3):569–582, 2015. 1, 2

work page 2015
[4]

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 4

work page 2009
[5]

K. He, X. Zhang, S. Ren, and S. Jian. Deep residual learn- ing for image recognition. In CVPR, 2016. 6, 10

work page 2016
[6]

Hou, M.-M

Q. Hou, M.-M. Cheng, X. Hu, A. Borji, Z. Tu, and P. Torr. Deeply supervised salient object detection with short con- nections. In CVPR, 2017. 5

work page 2017
[7]

P. Hu, B. Shuai, J. Liu, and G. Wang. Deep level sets for salient object detection. In CVPR, 2017. 3

work page 2017
[8]

L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE TPAMI, (11):1254–1259, 1998. 1

work page 1998

Show all 39 references

[9]

Jiang, J

H. Jiang, J. Wang, Z. Yuan, Y . Wu, N. Zheng, and S. Li. Salient object detection: A discriminative regional feature integration approach. In CVPR, 2013. 2, 6, 9

2013
[10]

D. A. Klein and S. Frintrop. Center-surround divergence of feature statistics for salient object detection. In ICCV,
[11]

Kr ¨ahenb¨uhl and V

P. Kr ¨ahenb¨uhl and V . Koltun. Efﬁcient inference in fully connected crfs with gaussian edge potentials. In NIPS
[12]

J. Kuen, Z. Wang, and G. Wang. Recurrent attentional networks for saliency detection. In CVPR, 2016. 3

2016
[13]

Li and Y

G. Li and Y . Yu. Visual saliency based on multiscale deep features. In CVPR, 2015. 3, 8

2015
[14]

X. Li, F. Yang, H. Cheng, W. Liu, and D. Shen. Contour knowledge transfer for salient object detection. In ECCV,
[15]

X. Li, L. Zhao, L. Wei, M.-H. Yang, F. Wu, Y . Zhuang, H. Ling, and J. Wang. Deepsaliency: Multi-task deep neural network model for salient object detection. TIP, 25(8):3919–3930, 2016. 3

2016
[16]

Y . Li, X. Hou, C. Koch, J. M. Rehg, and A. L. Yuille. The secrets of salient object segmentation. In CVPR, 2014. 8

2014
[17]

T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie. Feature pyramid networks for object detec- tion. In CVPR, 2017. 1, 2, 3, 4

2017
[18]

T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll´ar. Focal loss for dense object detection. In ICCV, 2017. 2, 5

2017
[19]

Liu and J

N. Liu and J. Han. Dhsnet: Deep hierarchical saliency network for salient object detection. In CVPR, 2016. 3

2016
[20]

N. Liu, J. Han, and M.-H. Yang. Picanet: Learning pixel- wise contextual attention for saliency detection. In CVPR,
[21]

J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 1, 3

2015
[22]

Z. Luo, A. Mishra, A. Achkar, J. Eichel, S. Li, and P.-M. Jodoin. Non-local deep features for salient object detec- tion. In CVPR, 2017. 2, 3, 6, 8, 9

2017
[23]

Manno-Kovacs

A. Manno-Kovacs. Direction selective contour detection for salient objects. TCSVT, 29(2):375–389, Feb 2019. 3

2019
[24]

Movahedi and J

V . Movahedi and J. H. Elder. Design and perceptual vali- dation of performance measures for salient object segmen- tation. In CVPR Workshops, 2010. 8

2010
[25]

Paszke, S

A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer. Automatic differentiation in pytorch. In NIPS Workshops,
[26]

Ronneberger, P

O. Ronneberger, P. Fischer, and T. Brox. U-net: Convo- lutional networks for biomedical image segmentation. In MICCAI, pages 234–241. Springer, 2015. 1, 3

2015
[27]

Simonyan and A

K. Simonyan and A. Zisserman. Very deep convolu- tional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 2, 4, 6, 8

2014 arXiv
[28]

Susstrunk, R

S. Susstrunk, R. Achanta, F. Estrada, and S. Hemami. Frequency-tuned salient region detection. In CVPR, 2009. 8

2009
[29]

L. Wang, H. Lu, Y . Wang, M. Feng, D. Wang, B. Yin, and X. Ruan. Learning to detect salient objects with image- level supervision. In CVPR, 2017. 7, 9

2017
[30]

L. Wang, L. Wang, H. Lu, P. Zhang, and X. Ruan. Saliency detection with recurrent fully convolutional networks. In ECCV, 2016. 3, 6, 9

2016
[31]

T. Wang, A. Borji, L. Zhang, P. Zhang, and H. Lu. A stagewise reﬁnement model for detecting salient objects in images. In ICCV, 2017. 6, 9

2017
[32]

T. Wang, L. Zhang, S. Wang, H. Lu, G. Yang, X. Ruan, and A. Borji. Detect globally, reﬁne locally: A novel approach to saliency detection. In CVPR, 2018. 2, 3, 6, 9, 10

2018
[33]

S. Woo, J. Park, J.-Y . Lee, and I. So Kweon. Cbam: Con- volutional block attention module. In ECCV, 2018. 7

2018
[34]

Q. Yan, L. Xu, J. Shi, and J. Jia. Hierarchical saliency detection. In CVPR, 2013. 2, 8

2013
[35]

C. Yang, L. Zhang, H. Lu, X. Ruan, and M.-H. Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013. 1, 3, 6, 8, 9

2013
[36]

Zhang, J

L. Zhang, J. Dai, H. Lu, Y . He, and G. Wang. A bi- directional message passing model for salient object de- tection. In CVPR, 2018. 6, 8, 9, 10

2018
[37]

Zhang, D

P. Zhang, D. Wang, H. Lu, H. Wang, and X. Ruan. Amulet: Aggregating multi-level convolutional features for salient object detection. In ICCV, 2017. 3, 6, 9 11

2017
[38]

Zhang, D

P. Zhang, D. Wang, H. Lu, H. Wang, and B. Yin. Learn- ing uncertain convolutional features for accurate saliency detection. In ICCV, 2017. 6, 9, 10

2017
[39]

Zhang, T

X. Zhang, T. Wang, J. Qi, H. Lu, and G. Wang. Progres- sive attention guided recurrent network for salient object detection. In CVPR, 2018. 2, 3, 6, 9, 10 12

2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Achanta, A

R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. S ¨usstrunk. Slic superpixels compared to state-of-the- art superpixel methods. IEEE TPAMI, 34(11):2274–2282,

work page

[2] [2]

S. Chen, X. Tan, B. Wang, and X. Hu. Reverse attention for salient object detection. In ECCV, 2018. 3, 6, 7, 9

work page 2018

[3] [3]

Cheng, N

M.-M. Cheng, N. J. Mitra, X. Huang, P. H. Torr, and S.-M. Hu. Global contrast based salient region detection. IEEE TPAMI, 37(3):569–582, 2015. 1, 2

work page 2015

[4] [4]

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 4

work page 2009

[5] [5]

K. He, X. Zhang, S. Ren, and S. Jian. Deep residual learn- ing for image recognition. In CVPR, 2016. 6, 10

work page 2016

[6] [6]

Hou, M.-M

Q. Hou, M.-M. Cheng, X. Hu, A. Borji, Z. Tu, and P. Torr. Deeply supervised salient object detection with short con- nections. In CVPR, 2017. 5

work page 2017

[7] [7]

P. Hu, B. Shuai, J. Liu, and G. Wang. Deep level sets for salient object detection. In CVPR, 2017. 3

work page 2017

[8] [8]

L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE TPAMI, (11):1254–1259, 1998. 1

work page 1998