A number of bottom-up saliency detection algorithms have been proposed in the literature. Since these have been developed from intuition and principles inspired by psychophysical studies of human vision, the theoretical relations among them are unclear. In this paper, we present a unifying perspective. Saliency of an image area is defined in terms of divergence between certain feature distributions estimated from the
central part and its surround. We show that various, seemingly different saliency estimation algorithms are in fact closely related. We also discuss some commonly
used center-surround selection strategies. Experiments with two datasets are presented to quantify the relative advantages of these algorithms.
Best student paper award in Computer Vision and Robotics Track
Learning Moving Cast Shadows for Foreground Detection (VS 2008)
Saliency Detection via Divergence Analysis: A Unified Perspective ICPR 2012
1. Saliency Detection via
Divergence Analysis
Jia-Bin Huang Narendra Ahuja
Electrical and Computer Engineering
University of Illinois, Urbana-Champaign
3. What this talk is about?
1. The saliency detection problem
2. A unifying framework for bottom-up
saliency detection algorithms
3. Ways to improve the performance
8. What is Saliency?
• Visual salience (or visual saliency) is the
distinct subjective perceptual quality
which makes some items in the world
stand out from their neighbors and
immediately grab our attention
Where to look?
15. Main Result
• Most of the bottom-up saliency detection
algorithms can be rewritten in the form of
divergence between probabilistic
distributions learned from center and
surround
34. Lessons Learned
• Most of the bottom-up saliency detection
algorithms are in fact closely related
– Not exhaustive, e.g., spectral-based methods
• How to improve the performance?
– Richer features
– Less approximation
– Adaptive center/surround support
37. Questions?
• Comments or questions:
Jia-Bin Huang
jbhuang1@illinois.edu
https://netfiles.uiuc.edu/jbhuang1/www/
38. References
• Saliency Detection via Divergence Analysis: An Unified Perspective
J.B. Huang and N. Ahuja. ICPR, 2012
• Frequency-tuned salient region detection
R. Achanta, S. Hemami, F. Estrada, and S. Susstrunk., CVPR 2009
• Saliency detection using maximum symmetric surround.
R. Achanta and S. Susstrunk, ICIP, 2010.
• State-of-the-art in visual attention modeling
A. Borji and L. Itti. PAMI 2012
• Saliency based on information maximization.
N. Bruce and J. Tsotsos. NIPS, 2005
• Global contrast based salient region detection.
M. M. Cheng, G. X. Zhang, N. J. Mitra, X. Huang, and S. M. Hu. CVPR, 2011
• The discriminant center-surround hypothesis for bottom-up saliency.
• D. Gao, V. Mahadevan, and N. Vasconcelos. NIPS 2007
39. References
• Context-aware saliency detection.
S. Goferman, L. Zelnik-Manor, and A. Tal. CVPR, 2010.
• Graph-based visual saliency.
J. Harel, C. Koch, and P. Perona. NIPS, 2006
• Saliency detection: A spectral residual approach.
X. Hou and L. Zhang. CVPR, 2007
• A model of saliency based visual attention for rapid scene analysis.
L. Itti, C. Koch, and E. Niebur. PAMI 1998
• Learning to predict where humans look.
T. Judd, K. Ehinger, F. Durand, and A. Torralba. ICCV, 2009
• Center-surround divergence of feature statistics for salient object detection.
D. A. Klein and S. Frintrop. ICCV, 2011
• Saliency detection based on frequency and spatial domain analyses
J. Li, M. Levine, X. An, and H. He. BMVC 2011
40. References
• Learning to detect a salient object
T. Liu, J. Sun, N. N. Zheng, X. Tang, and H. Y. Shum, CVPR 2007
• Contrast-based image attention analysis by using fuzzy growing
Y. Ma and H. Zhang, ACM MM 2003
• Top-down control of visual attention in object detection
A. Oliva, A. Torralba, M. Castelhano, and J. Henderson. ICIP 2003
• Segmenting salient objects from images and videos
E. Rahtu, J. Kannala, M. Salo, and J. Heikkil., ECCV 2010
• Computational versus psychophysical image saliency: A comparative
evaluation study, A. Toet. PAMI 2011
• Visual attention detection in video sequences using spatiotemporal cues
Y. Zhai and M. Shah. ACM MM 2006
• Sun: A bayesian framework for saliency using natural statistics
L. Zhang, M. Tong, T. Marks, H. Shan, and G. Cottrell. JOV 2008
Notas del editor
Thanks for the introduction. I am Jia-Bin Huang. Today I am going to talk about some interesting facts about saliency detection.
This is a joint work with Prof. Narendra Ahuja at the University of Illinois, Urbana-Champaign
I will start with an overview of the problem. Then, introducing a unifying framework for bottom-up saliency detection algorithms. Later, we suggest ways to improve the performance.
Let’s play a game first. Please clap when you see something different.
Saliency is the distinct subjective quality which makes some items pop out from their neighbors and grab our attention. It guides us where to look.
Here are some examples of stimulus.
The reasons that this problem is important because it can facilitate numerous applications in computer vision and computer graphics.
Here is the problem setting. Given an input image, our goal is to produce a saliency map locating the salient regions.
There have been many design principles that guide the development of saliency detection algorithms. For example, rarity, complexity, contrast, spectral methods, and learning-based approaches.
However, since some of the principles are inspired from psychological experiments, some are from intuitions, and others are purely computational, it’s often difficult to see their relations and more practically, which one is better?
In this paper, we provide the bridges between these methods. Our claim is that most of the bottom-up saliency detection algorithms can be rewritten in the form of divergence between the probabilistic distributions learned from center and surround.
Let’s start with KL-divergence. KL divergence is a contrast function between two probabilistic distributions. It has the following forms. It has many important operational meanings in detection, estimation, and information theory. In the following few slides, I will explicitly show the connections between various models.
If we assume that the center support is a single pixel x_i, the KL-divergence from center to surround will become the so-called Shannon’s self-information. This class of saliency is called rarity-based saliency.
If we take the difference between the self-information of center and surround, it yields the saliency in Rahtu’s paper at ECCV. If we make the assumption of feature channel independence, we can derive the saliency measure in Klein at ICCV.
So how about switch the order? From the likelihood theory we know that the divergence can be rewritten as likelihood function. This forms the basics of contrast-based saliency methods.
Assume that the distribution of the center follows Laplacian distributions. The saliency measure has this form.
Similarly, assuming that the distribution is Gaussian. Using various approximation approaches gives rises to various measures.
Now let’s look at some symmetrised divergence measure. The first one is the symmetric KL divergence. This measure has been widely used from feature selection. In the context of saliency detection, it corresponds to combine local and global rarity. Lambda divergence is another symmetric measure. It corresponds to the mutual information-based saliency.