600 likes | 756 Views
Context-aware saliency. Stas Goferman Lihi Zelnik -Manor Ayellet Tal. What is saliency?. …. Please describe this picture. Picture description. Man in a flower field In the fields Spring blossom. Please describe this picture. Picture description. Olympic weight lifter
E N D
Context-aware saliency StasGoferman LihiZelnik-Manor Ayellet Tal
Picture description Man in a flower field In the fields Spring blossom
Picture description Olympic weight lifter Olympic victory Olympic achievement
Segmentation • Man in a flower field • In the fields • Spring blossom • Olympic weight lifter • Olympic victory • Olympic achievement
Context-aware salient regions • Man in a flower field • In the fields • Spring blossom • Olympic weight lifter • Olympic victory • Olympic achievement
Principles of saliency Following perceptual properties
Principle 1 • Local low-level factors • Contrast • Color
Walther & Koch – local filtering [Walther and Koch, Neural Networks 2006]
Principle 1 Walther & Koch, 2006 • Local low-level factors • Contrast • Color
Principle 2 • Global considerations • Maintain unique features
Hou & Zhang “spectral residual” [Hou & Zhang CVPR 2007]
Principle 2 Hou & Zhang, 2007 • Global considerations • Maintain unique features
Principles 1 + 2 Local & global
Liu et al – learning saliency Input Multi-scale contrast Center surround Color Final [Liu et al, CVPR 2007]
Principles 1 + 2 Liu et al, 2007 Local & global
Principle 3 • Visual organization (Gestalt) • Few centers of gravity [Koffka] • Position is important!!
Principle 4 Low-level With face detection [Judd et al, ICCV 2009] • High-level • Faces • Objects • People • …
Incorporating the 4 Principles Our result
Global Hou & Zhang, 2007 Local Walther & Koch, 2006 Local + global Liu et al, 2007 Our result
The “how” The steps of our algorithm
Local-global saliency Not salient salient Principles 1-2: Unique appearance salient
Computing uniqueness Principles 1-2: Unique appearance salient
Appearance uniqueness Euclidean distance between colors of patches at pi & pj Principles 1-2: Unique appearance salient
Appearance uniqueness high salient Principles 1-2: Unique appearance salient
Positional information Similar patches both near and far Not salient Principle 3: Position is important!
Positional information Similar patches near Salient Principle 3: Position is important!
Positional information Normalized Euclidean distance between positions of pi & pj Principle 3: Position is important!
Single scale uniqueness High salient Distance between a pair of patches:
Single scale uniqueness High for K most similar salient Distance between a pair of patches:
Single scale saliency K most similar patches at scale r
Multiple scales Scale 1 Scale 4 • Salient at: • Multiple scales foreground • Few scales background
Including immediate context Context • Principle 3: • Few centers of gravity
Visual organization Focus points Distance map Final result X
Algorithm summary X Single-scale saliency Multiple scales Final saliency
Results …
Non-interesting background Our result Walther & Koch, 2006 Hou & Zhang, 2007
Non-interesting background Our result Walther & Koch, 2006 Hou & Zhang, 2007
Object + immediate surrounding Our result Walther & Koch, 2006 Hou & Zhang, 2007
Object + immediate surrounding Our result Walther & Koch, 2006 Hou & Zhang, 2007
Complex scenes Our result Walther & Koch, 2006 Hou & Zhang, 2007
Complex scenes Our result Walther & Koch, 2006 Hou & Zhang, 2007
Quantitative evaluation Database of Hou & Zhang
Judd et al. database Our + center Judd Our
Judd et al. database Our + center Judd Our
Boiman & Irani[IJCV’07] Input Boiman & Irani Our result