300 likes | 493 Views
Efficient Salient Region Detection with Soft Image Abstraction Presented by: Shai Krakovski. A Work of: Ming-Ming Chen, Jonathan Warrell , Wen -Yan Lin, Shuai Zheng , Vibhav Vineet , Nigel Crook Vision Group, Oxford Brookes University. Lecture structure. Motivation Related work
E N D
Efficient Salient Region Detection with Soft Image AbstractionPresented by: ShaiKrakovski A Work of: Ming-Ming Chen, Jonathan Warrell, Wen-Yan Lin, ShuaiZheng, VibhavVineet, Nigel Crook Vision Group, Oxford Brookes University
Lecture structure • Motivation • Related work • Proposed solution • Results • Pros & Cons. • What can be done?
Motivation • Object of interest image segmentation • Adaptive compression • Object level image manipulation • Internet visual media retrieval
Motivation • Object of interest image segmentation • Adaptive compression • Object level image manipulation • Internet visual media retrieval
Motivation • Object of interest image segmentation • Adaptive compression • Object level image manipulation • Internet visual media retrieval
Motivation • Object of interest image segmentation • Adaptive compression • Object level image manipulation • Internet visual media retrieval
Gofferman, Zelnik-Manor& Tal 2010 • Multiscale Patch-match • Normalization by dominance proximity • Higher levelsegmentation
Chang, Zhang et al 2011 • Histogram-based Contrast • Region-based Contrast
Perazzi &Krahenbuhl2012 • Superpixels by K-means • Uniqueness • Distribution • .
Image Abstraction by GMM Input GMM Output Correlation
Gaussian Mixture Models e1 mean
Image Abstraction by GMM Input GMM Output Correlation
Correlation Output
Global Cues • Global Uniqueness (GU) • Color Spatial Distribution (CSD)
Global Cues • Global Uniqueness (GU) • Color Spatial Distribution (CSD)
Global Cues • Global Uniqueness (GU) • Color Spatial Distribution (CSD)
Pros & Cons. Method • Cons • Only considers colors • Pros • Fast • Good results Paper • Pros • Organized • Many images • Time • Cons • Few explanations • Limitations • Nothing new
What can be done? • Pre-processing of the 3 channels. • Better integration of GU and CSD. • Comparing the GMM with SVM and K-means. • Finding other\better Global Cues. • Useful application.