Does saliency mean importance?

It usually means prominence under a specified model; importance depends on the task and context.

Is a saliency score a probability?

Only if the method defines and validates that interpretation.

Are explanation heatmaps the same as visual-attention maps?

No. They can represent different quantities and require different validation.

AI Glossary / Computer vision and multimodal AI / Advanced

Saliency

Definition

In visual perception and computer vision, saliency describes how strongly a location or feature stands out within a scene under a stated model of attention. A saliency map assigns spatial scores representing that prominence. The term can also refer to model-explanation maps, so the intended meaning must be specified rather than assuming every saliency map measures human attention.

What it means

A brightly contrasting sign can attract attention even when it is irrelevant to a viewer's task. Conversely, someone searching for a particular object may examine a visually unobtrusive region. Saliency therefore concerns prominence under particular conditions, not a universal ranking of importance. A computational map is an estimate produced by a method, not a direct reading of every viewer's thoughts. Its usefulness depends on the scene, the viewing task and what the scores are designed to represent.

How it works

A classic bottom-up approach compares visual features across locations and scales, then combines feature contrasts into a map. The influential Itti - Koch - Niebur model uses intensity, color and orientation information to guide location selection. Other approaches learn spatial predictions from data. Outputs require interpretation: a relative prominence score need not be a calibrated probability of fixation. Task-driven attention introduces additional influences. In model interpretation, a saliency map instead may describe how a prediction responds to input features; that computational sensitivity is not automatically the same as perceptual conspicuity.

Three practical examples

A design team uses a visual saliency estimate to identify distracting regions in a draft layout before conducting a user study.
An image-processing exercise compares maps for an isolated colored shape and a uniform field of similar shapes.
A research team evaluates whether a fixation-prediction model matches observed gaze under a clearly defined viewing task.

Think of it this way

Imagine entering a room where one lamp flashes while everything else is still. The lamp may draw your gaze, but you might immediately look elsewhere if you are searching for your keys. Saliency resembles the pull of the conspicuous lamp, while your search goal adds another influence. The analogy shows why attention-grabbing features and task-relevant information can overlap without being identical, and why a prominence map cannot fully describe a person's priorities.

Worked example

Consider an original toy map with three regions assigned nonnegative scores 2,6 and 2. Their total is 10, so dividing by the total produces normalized weights 0.2,0.6 and 0.2. Region two receives the largest relative weight under this map. However, normalization alone does not establish that 60% of viewers will look there, or that it will receive 60% of viewing time. Those claims would require an appropriate probabilistic model and validation against observations. The example demonstrates score normalization and ranking, not a complete saliency algorithm or a behavioral prediction.

Illustrative case study

A fictional educational website tests a page containing a lesson title, an illustration and a brightly colored promotional panel. A saliency tool emphasizes the panel, prompting the team to investigate whether it distracts learners. In a small task-based evaluation, users still find the lesson title, but several overlook supporting instructions. The team revises the layout and measures task completion again. This scenario shows saliency as one diagnostic input. It does not equate a heatmap with comprehension, learning, commercial success or an exhaustive account of how users read the page.

Common misconceptions and mistakes

The most salient region is not necessarily the most meaningful, useful or memorable. Another mistake is interpreting every colored heatmap as a probability distribution. Relative scores, fixation densities and prediction-sensitivity maps have different meanings. A map computed for one image size or viewing condition may change under another. Center preferences in a dataset can also affect predictions. Finally, strong agreement with a benchmark does not establish that a method explains all human attention, especially when real viewers have goals that differ from the benchmark's viewing instructions.

Business use

Saliency estimates can help prioritize visual review, compare draft layouts or allocate processing attention. Define the intended use before choosing a tool: predicting gaze, identifying prominent objects and explaining a classifier are different tasks. Validate against relevant observations, such as gaze measurements or user task performance, rather than judging only whether a heatmap looks plausible. Combine visual prominence with readability and accessibility assessment. Keep the results in context so teams do not optimize for attention-grabbing elements at the expense of helping users understand and complete their intended tasks.

Check your understanding

Which toy region ranked highest?

Answer: Region two, with score 6.

What did its normalized weight of 0.6 establish?

Answer: Its share of the stated score total.

Why did it not prove a 60% fixation rate?

Answer: Normalizing arbitrary scores does not by itself create a calibrated behavioral probability model.

Related terms

Scanpath
Visual attention
fixation prediction
conspicuity map
salient-object detection
feature attribution

Learn next

Compare saliency maps for the same scene under different contrast arrangements. Then introduce a specific search task and consider which locations a viewer needs to inspect. Label each map's output quantity explicitly before comparing it with human observations or model behavior.

Sources

Itti, Koch and Niebur, A Model of Saliency-Based Visual Attention for Rapid Scene Analysis: https://hasler.ece.gatech.edu/Courses/MachineLearning/FoundationalPapers/Itti_Koch_Neiber1998.pdf The three-region calculation, flashing-lamp analogy and educational-website case are original illustrations. The historical model provides a concrete example of bottom-up saliency rather than a universal definition of every modern attention or attribution method.

Previous: Safety Next: Salt-and-pepper noise