Global Contrast based Salient Region Detection MingMing Cheng

Global Contrast based Salient Region Detection Ming-Ming Cheng, Niloy J. Mitra, Xiaolei Huang, Philip H. S. Torr, and Shi-Min Hu speaker: Jianhong Yuan date: 2017/6/20 1

salient object and saliency 2

related work • Y. Zhai and M. Shah, 2006; luminance information • R. Achanta, S. Hemami, F. Estrada, and S. Susstrunk, 2009; average image color (a) original (b)Zhai and Shah (c)Achanta 3

Key work Histogram-based contrast(HC) Region-based contrast(RC) (a) original (b) HC map (c) RC map 4

high contrast spatial relationship Human attention HC: high contrast RC: high contrast ; spatial relationship ; improvement of HC 5

HC define saliency value color distance matric time complexity: O(N)+O(n 2) the number of pixel colors in the image 6

Histogram speed up (a) input image (b) color histogram of (a) (c) quantized image 7

saliency value Color space smooth before after Redefine saliency value of each color defined by the weighted average of the saliency values of similar colors 8

improvement of HC pixel-level contrast spatial relationship high computational cost region-level contrast region contrast(RC) 9

Algorithm flow (RC) image segment create color histogram initial saliency value refinement image regions RC-map 10

Saliency value of region • define the initial saliency for each region HC: RC: color distance matric between the two pixels color distance matric between the two regions weight of the region ri. 11

Refinement of RC Spatial weighted region contrast (initial saliency) spatial prior weighting term 12

Application Saliency. Cut Sketch based image retrival 13

Grab. Cut (a) input (b) output 14

Saliency. Cut candidate region background Saliency. Cut 15

Saliency. Cut Segmentation by iterative fitting original initial segment trimap foreground background second trimap red --> foreground final segment manual labeled green --> background 16

Sketch based image retrival Salient object outlines input sketch comparing consistency Shape-based sorting retrival result 17

Sketch based image retrival Sketch based image comparison 18

Evaluation segmentation by fixed thresholding object of interest image segmatation 19

Segmentation by fixed thresholding (i) Achanta et al. dataset (ii) MSRA 10 K dataset 20

Object of interest image segmatation (i) Achanta et al. dataset (ii) MSRA 10 K dataset 21

Conclusion • present global contrast based saliency computation methods —— Histogram based contrast(HC) • spatial information-enhanced Region based Contrast (RC). • introduce a novel unsupervised segmentation algorithm —— Saliency. Cut 22

limitations • images with multiple objects may not get all the salient objects. • if the objects occlude each other, cannot generate good results. 23

Thank you 24

questions Yanghao • Why choose lab color space instead of RGB color model? Lab color space is based on our physiological characteristics, and contain more wider color gamut. 25

questions Xie kaixi • while the quantification based on RGB model? Actually, the author tried to quantify the RGB and LAB space, choose the effect of a better one 26

questions An weizhi You mentioned that the method you are using does not require an iterative process compared with Grab. Cut? I mean compared to Grab. Cut, we don't need any manual annotation to segment background region, it's a iterative process of grab. Cut. How to segment image into regions? We use the fixed threshold to divide image into two parts 27