In The Name of God Salient Edges A

In The Name of God Salient Edges: A Multi-scale Approach M. Holtzman-Gazit, L. Zelnik-Manor and I. Yavne, ECCV 2010 Workshop on Vision for Cognitive Tasks. Presented by Saba Aflaki Ghazaal Haeri May 31, 2011

Outline Key words Objective Methods The Algorithm Conclusion References 12/1/2020 2

Visual Salience �Distinctive subjective perceptual quality �Make some items in the world stand out from their neighbors. �Grab our attention immediately. 12/1/2020 3

Visual Salience �Top-down Bottom-up �User-driven Stimulus-driven �High-level Low-level info. . �� Color, Content Intensity, semantics: Orientation, … good weather, cloud, … 12/1/2020 4

Experiencing Visual Salience �Immediate attention �No scanning Orientation pop-out Color pop-out Not locally determined! 12/1/2020 5

Experiencing Visual Salience Distinct motion Natural visual salience pattern 12/1/2020 6

Scale-space �Handling image structures at different scales. �Representing an image as a one-parameter family of smoothed images. �Scale parameter t �Object size � largely 12/1/2020 smoothed away 7

Linear Gaussian Representation f(x, y) 12/1/2020 f(x, y) g(x, y; t) t=256 t=64 t=16 t=4 t=1 t=0 L(x, y, t) 8

Beltrami Representation Edge-Preserving Smoothing 12/1/2020 9

Next… Key words Objective Methods The Algorithm Conclusion References 12/1/2020 10

Objective �Recovery of salient objects’ boundaries • Object tracking • Video encoding • Automatic retargeting • Seam carving for content -aware image resizing 12/1/2020 11

Next… Key words Objective Methods The Algorithm Conclusion References 12/1/2020 12

Saliency Estimation Methods �Single-scale �Part of larger textures appear as salient because of locally high-contrast changes �Multi-scale �Combine different edge information from different levels 12/1/2020 13

Saliency Estimation Methods �Global �Coarse, blurry map Global �Lack fine detail Local �Miss parts of salient objects �Local �Detect background texture as salient 12/1/2020 14

Next… Key words Objective Methods The Algorithm Conclusion References 12/1/2020 15

Multi-scale Bottom-up Approach Local Saliency Salient Edges Regional Saliency 12/1/2020 Saliency Map 16

Single-level Phase Based Saliency Local Saliency 12/1/2020 17

Single-level Phase Based Saliency 12/1/2020 18

Single-level Phase Based Saliency g Gaussian filter 12/1/2020 19

Beltrami vs. Gaussian After. Saliency Gaussian. Map filtering 12/1/2020 After. Saliency Beltrami. Map filtering 20

Multi-level Phase Based Saliency �Increase the saliency of pixels which are salient at multiple levels (use harmonic mean). �Eliminate occasionally isolated pixels which are detected as salient (use 5 5 median filter). 12/1/2020 21

Multi-level Local Saliency Not Enough! 12/1/2020 22

Single-level Regional Saliency �Divide image into a grid of blocks of size ki �Compute for each block the RGB color histogram (48 bins, 16 for each color). Regional Saliency �Compare color histogram of blocks. �Use 2 distance �blocks in the neighborhood of size m m 12/1/2020 23

Single-level Regional Saliency C=3 x normalized horizontal distance y normalized vertical distance d. H(Bk, Bl) 12/1/2020 Similarity 24

Single-level Regional Saliency M 1/6 of the number of surrounding blocks in the m m neighborhood m=3 d. H(Bk, Bl) 12/1/2020 Sr(Bk) Saliency 25

Single-level Regional Saliency �Saliency maps smaller by a factor of �Re-scale to the original size with bilinear filter 12/1/2020 26

Multi-level Regional Saliency �Areas with high regional saliency value in all levels are considered as salient (use geometric mean). 12/1/2020 27

Multi-level Regional Saliency 12/1/2020 28

Combining Regional and Local Saliency Sregional Slocal Salient with respect to neighborhood Important Edges ST =Slocal Important edges that define the salient objects 12/1/2020 29

Result 12/1/2020 30

Next… Key words Objective Methods The Algorithm Conclusion References 12/1/2020 31

Conclusion �Novel method for saliency detection �Integrates local and regional saliency at multiple scales �does not favor one level over another �low computational cost O(MN) �works well on highly textured images 12/1/2020 32

Conclusion Table 1. ROC areas for different saliency models with respect to all human eye fixation or human labeling in the data sets of [16] and [5]. n = 4 levels for local saliency k=[8, 16, 32, 64, 128] for regional saliency 12/1/2020 33

Next… Key words Objective Methods The Algorithm Conclusion References 12/1/2020 38

References � M. Holtzman-Gazit, L. Zelnik-Manor and I. Yavne , Salient Edges: A Multi Scale Approach, ECCV 2010 Workshop on Vision for Cognitive Tasks � Guo, C. , Ma, Q. , Zhang, L. : Spatio-temporal saliency detection using phase spectrum of quaternion fourier transform, CVPR (2008). � Patrick Bas, Nicolas Le Bihan and Jean-Marc Chassery, Color Image Watermarking Using Quaternion Fourier Transform, Multimedia and Expo, 2008 IEEE International Conference. � http: //www. scholarpedia. org/article/Visual_salience � http: //en. wikipedia. org/wiki/Scale_space � http: //en. wikipedia. org/wiki/Image_resolution#Pixel_resolution � http: //www. cs. technion. ac. il/~ron/belt- html/node 1. html#SECTION 000100000000 12/1/2020 35

Thank you! 12/1/2020 Any Questions? 36