Supplementary Material II: Additional Experiment Results

advertisement
Supplementary Material II: Additional Experiment Results
This document provides the following additional experiment results for the paper Minimum Barrier Salient
Object Detection at 80 FPS. The evaluated methods are FT [1], HC [2], SIA [3], RC [2], GS [6], HSal [7], AMC
[4], SO [8] and our methods MB and MB+, together with the baseline method GD using geodesic distance.
1. The AUC (Area Under the Curve) scores of all the models (Fig. 1).
2. The mAE (mean Absolute Error) scores of all the models (Fig. 2).
3. The precision rate, recall rate and F-measure scores using the adaptive thresholding method [1] (Fig. 3).
4. The weighted-Fβ scores when the proposed post-processing function is applied and optimized for all the
compared models (Fig. 4). In our post-processing function, there are three components, 1) smoothing 2)
centeredness and 3) contrast enhancement. We try all possible 23 = 8 combinations of these components
(including none of the three and all of the three), and report the best score for each model.
5. More sample saliency maps (Figs. 5-8).
References
[1] R. Achanta, S. Hemami, F. Estrada, and S. Susstrunk. Frequency-tuned salient region detection. In CVPR, 2009.
[2] M.-M. Cheng, N. J. Mitra, X. Huang, P. H. S. Torr, and S.-M. Hu. Global contrast based salient region detection. TPAMI,
37(3):569–582, 2015.
[3] M.-M. Cheng, J. Warrell, W.-Y. Lin, S. Zheng, V. Vineet, and N. Crook. Efficient salient region detection with soft image
abstraction. In CVPR, 2013.
[4] B. Jiang, L. Zhang, H. Lu, C. Yang, and M.-H. Yang. Saliency detection via absorbing markov chain. In ICCV, 2013.
[5] R. Margolin, A. Tal, and L. Zelnik-Manor. What makes a patch distinct? In CVPR, 2013.
[6] Y. Wei, F. Wen, W. Zhu, and J. Sun. Geodesic saliency using background priors. In ECCV. 2012.
[7] Q. Yan, L. Xu, J. Shi, and J. Jia. Hierarchical saliency detection. In CVPR, 2013.
[8] W. Zhu, S. Liang, Y. Wei, and J. Sun. Saliency optimization from robust background detection. In CVPR, 2014.
1
1
THUS10000
DUTOmron
ECSSD
PASCAL-S
0.95
0.9
0.95
0.9
0.9
0.85
0.85
0.8
0.8
0.85
0.8
0.75
0.75
0.8
AUC
0.85
AUC
AUC
AUC
0.9
0.7
0.7
0.7
0.65
0.65
0.6
0.65
0.75
0.75
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
Figure 1: AUC scores of the compared methods. The AUC metric is used previous works like [5].
DUTOmron
THUS10000
ECSSD
PASCAL-S
0.4
0.35
0.35
0.25
0.35
0.3
0.3
0.2
0.3
0.15
mAE
mAE
mAE
mAE
0.25
0.25
0.2
0.25
0.2
0.15
0.1
0.2
0.15
0.1
0.15
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
Figure 2: mAE (mean Absolute Error) scores of the compared methods. The mAE metric is used previous works
like [3]. Lower scores under this metric indicate better performance.
THUS10000
0.8
0.9
0.75
0.85
0.7
0.8
0.65
0.75
0.6
ECSSD
DUTOmron
precision
recall
Fβ
PASCAL-S
0.75
0.8
0.7
0.75
0.65
0.7
0.6
0.65
0.55
0.6
0.7
0.55
0.65
0.5
0.6
0.45
0.45
0.4
0.55
0.4
0.4
0.35
0.35
0.35
0.3
0.3
0.5
0.45
0.55
0.5
0.5
0.45
0.3
0.25
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
Figure 3: Precision, recall and F-measure under the adaptive thresholding setting. As in [1], a gray-scale saliency
map is thresholded at 2 times of its mean saliency value. The resultant segmentation mask is evaluated based on
the ground truth mask. As in [1], we use β 2 = 0.3 for the F-measure calculation.
MSRA10K
ECSSD
PASCAL−S
0.55
0.4
0.5
0.4
0.3
0.35
0.3
0.25
0.55
0.5
0.5
0.45
β
0.6
0.45
Weighted−F
0.45
Weighted−F β
0.7
Weighted−F β
0.6
β
Weighted−F
DUTOmron
0.4
0.35
0.3
0.2
0.15
0.3
0.25
0.25
0.2
0.4
0.35
0.2
0.2
0.15
0.15
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
(a)
MSRA10K
ECSSD
PASCAL−S
0.55
0.4
0.5
0.4
0.3
0.35
0.3
0.25
0.55
0.5
0.5
0.45
β
0.6
0.45
Weighted−F
0.45
Weighted−F β
0.7
Weighted−F β
0.6
β
Weighted−F
DUTOmron
0.4
0.35
0.3
0.2
0.15
0.3
0.25
0.25
0.2
0.4
0.35
0.2
0.2
0.15
0.15
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
B+
M
B
M
GD
SO C
AM
HS
GS
RC
SIA
HC
FT
(b)
Figure 4: Weighted-Fβ scores when the effect of post-processing is controlled. In (a), we show the original scores
for the compared methods. In (b), we show the scores after controlling the factor of post-processing (see text for
more details). The post-processing improves the score of FT and HC, but does not significantly improve or even
degrade the scores of the other compared models. As we can see, controlling this post-processing factor does not
change the rankings of our methods.
Input
FT
HC
SIA
RC
GS
HS
AMC
SO
GD
Figure 5: Sample saliency maps of the compared models.
MB
MB+
GT
Input
FT
HC
SIA
RC
GS
HS
AMC
SO
GD
Figure 6: Sample saliency maps of the compared models.
MB
MB+
GT
Input
FT
HC
SIA
RC
GS
HS
AMC
SO
GD
Figure 7: Sample saliency maps of the compared models.
MB
MB+
GT
Input
FT
HC
SIA
RC
GS
HS
AMC
SO
GD
Figure 8: Sample saliency maps of the compared models.
MB
MB+
GT
Download