Saliency Map Analysis for Interface Attention Prediction

Description

Uses computational saliency models to predict which regions of a screen will attract early visual attention, checking designs before eye tracking or live testing.

Saliency models compute where bottom-up visual features such as colour contrast, intensity, orientation and, in deep models, learned features like faces and text, will draw early attention. Itti, Koch and Niebur's model combines feature maps into a saliency map; deep models such as DeepGaze predict human fixations much more accurately on natural images, and models trained on webpages and UIs perform better still for interfaces. In UX, saliency maps provide quick checks: does the primary call to action stand out, are warnings competing with decoration, is visual hierarchy consistent with task priority? Rigour requires understanding the limits: models predict early, stimulus-driven attention, not task-driven search or comprehension, so predictions should be validated against eye tracking or click data and used comparatively between variants rather than as absolute truth.

Attributes

Loss type
Unmigrated
Stage
Design
Effort
Hours

Steps

  1. Apply a saliency model (open-source: MIT Saliency Benchmark; Python: pySaliency) to interface screenshots. Compute predicted fixation density map. Compare to:
  2. designer's intended attention hierarchy (what should users look at first?),
  3. actual eye-tracking data (if available). Overlap coefficient between predicted and intended: < 0.60 = visual hierarchy redesign needed. Use as a rapid pre-testing filter.

Sources

  • Computer Vision: A Modern Approach — David A. Forsyth
  • Itti, Koch & Niebur, A model of saliency-based visual attention for rapid scene analysis (IEEE PAMI, 1998)
  • Kümmerer et al., DeepGaze II (2016)

Also known as

  • Visual saliency modelling
  • Predicted attention heatmaps
  • Computational attention prediction

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning