What is a saliency map?

It is an image-sized map that scores how likely each location is to attract a viewer's gaze, with bright spots marking the regions predicted to stand out most.

Why is predicting attention useful?

Knowing where people look lets systems concentrate rendering quality, compression bits, or design emphasis on the regions that matter most to viewers, saving effort where attention is unlikely to fall.

Visual Saliency and Attention

Visual saliency and attention concern where in an image people are most likely to look, and computational models predict this to guide graphics, vision, and interface design.

Definition

Visual saliency is the property that makes some image regions stand out and attract gaze, and saliency modeling is the computational prediction of where attention and fixations will fall.

Scope

This topic covers bottom-up saliency driven by contrast in features such as intensity, color, and orientation, top-down attention guided by tasks and goals, the prediction of human gaze and fixations, and applications that direct rendering, compression, and design toward attended regions.

Core questions

What makes a region of an image attract attention?
How are bottom-up and top-down influences on attention combined?
How accurately can human gaze be predicted from an image?
How can saliency guide graphics and vision systems?

Key concepts

Saliency maps
Center-surround contrast
Feature integration
Bottom-up and top-down attention
Fixation and gaze prediction
Attention benchmarks

Key theories

Feature-integration saliency model: Saliency is computed by extracting feature maps for intensity, color, and orientation, detecting local center-surround contrast in each, and combining them into a master map whose peaks predict where attention is drawn.
Bottom-up versus top-down attention: Attention is steered both by stimulus-driven salience and by task-driven goals, and computational models increasingly integrate both, a distinction central to evaluating and improving gaze prediction.

Clinical relevance

Saliency models guide perceptually driven rendering and compression that allocate effort to attended regions, inform user-interface and advertising design, support automatic image cropping and retargeting, and contribute to robotics and assistive vision.

History

Grounded in psychological theories of attention, the Itti-Koch-Niebur model of 1998 gave an influential computational account of bottom-up saliency; benchmarks and surveys consolidated the field, and deep networks later improved gaze prediction substantially.

Key figures

Laurent Itti
Christof Koch
Ali Borji

Seminal works

itti1998
borji2013

Frequently asked questions

What is a saliency map?: It is an image-sized map that scores how likely each location is to attract a viewer's gaze, with bright spots marking the regions predicted to stand out most.
Why is predicting attention useful?: Knowing where people look lets systems concentrate rendering quality, compression bits, or design emphasis on the regions that matter most to viewers, saving effort where attention is unlikely to fall.