Saliency model application

Description

Applies computational models of bottom-up visual saliency to screens to predict where attention will be drawn first, before any eye-tracking study is run.

Saliency models compute, for each point in an image, how much it stands out in colour, intensity, orientation and other features relative to its surroundings, combining them into a saliency map. Itti and Koch's model was biologically inspired; modern deep models trained on eye-tracking data predict fixations considerably better. Applied to interfaces, they indicate whether critical elements are salient and whether decoration competes with them. Commercial 'predictive eye-tracking' tools are mostly renamed saliency models, often presented without their limitations: they capture bottom-up salience but not task-driven, top-down attention, which dominates in expert work. Rigour comes from using models validated on UI images (for example models trained on webpage datasets), comparing salience ranks with intended priority, and validating with real eye tracking where stakes are high.

Attributes

Loss type
Renamed
Stage
Design
Effort
Hours

Steps

  1. Capture screenshots of key screens in realistic states, including alerts and busy conditions.
  2. Run at least one validated saliency model, preferably one trained on UI or webpage data, to produce saliency maps.
  3. Define areas of interest for critical elements and compute their mean saliency rank relative to other elements.
  4. Flag critical elements outside the top 3 most salient regions and decorative elements that rank in the top 3.
  5. Adjust the design and re-run the model to compare versions.
  6. For safety-critical screens, validate with a short eye-tracking study (10-15 participants) measuring time to first fixation on critical areas.

Sources

  • Bear, M., Connors, B. & Paradiso, M., Neuroscience: Exploring the Brain
  • Garrett, B., Brain & Behavior: An Introduction to Behavioral Neuroscience
  • Gopinath, V., Methods Atlas
  • Neuroscience (Bear et al.)
  • Brain & Behaviour (Garrett)
  • Methods Atlas (Gopinath)
  • The Inquiry Series, Ch 9 — Attention — The Gatekeeper

Also known as

  • Computational saliency modelling
  • Visual saliency prediction
  • Attention heat-map prediction
  • Visual saliency modelling
  • Predicted attention heatmaps
  • Computational attention prediction
  • Saliency Map Analysis for Interface Attention Prediction

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning