Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Impact of Design Decisions in Scanpath Modeling

Parvin Emami, Yue Jiang, Zixin Guo, Luis A. Leiva

TL;DR

This work addresses the sensitivity of GUI scanpath predictions to non-learned design parameters in saliency models. By systematically varying input image size, IOR decay, and masking radius in DeepGaze++ on the UEyes dataset, the authors quantify impacts across DTW, Eyenalysis, Determinism, and Laminarity, and demonstrate the generalizability of findings to other models and datasets like MASSVIS. They introduce a generalized IOR decay, $\gamma^{(n - i - 1)}$, and identify robust parameter settings (e.g., square inputs, $\gamma = 0.1$, masking radius around $0.05$) that improve several metrics. The results offer practical guidance for building and evaluating scanpath models and underscore the importance of reporting design choices in GUI saliency research, with publicly available code and models to promote replicability.

Abstract

Modeling visual saliency in graphical user interfaces (GUIs) allows to understand how people perceive GUI designs and what elements attract their attention. One aspect that is often overlooked is the fact that computational models depend on a series of design parameters that are not straightforward to decide. We systematically analyze how different design parameters affect scanpath evaluation metrics using a state-of-the-art computational model (DeepGaze++). We particularly focus on three design parameters: input image size, inhibition-of-return decay, and masking radius. We show that even small variations of these design parameters have a noticeable impact on standard evaluation metrics such as DTW or Eyenalysis. These effects also occur in other scanpath models, such as UMSS and ScanGAN, and in other datasets such as MASSVIS. Taken together, our results put forward the impact of design decisions for predicting users' viewing behavior on GUIs.

Impact of Design Decisions in Scanpath Modeling

TL;DR

This work addresses the sensitivity of GUI scanpath predictions to non-learned design parameters in saliency models. By systematically varying input image size, IOR decay, and masking radius in DeepGaze++ on the UEyes dataset, the authors quantify impacts across DTW, Eyenalysis, Determinism, and Laminarity, and demonstrate the generalizability of findings to other models and datasets like MASSVIS. They introduce a generalized IOR decay, , and identify robust parameter settings (e.g., square inputs, , masking radius around ) that improve several metrics. The results offer practical guidance for building and evaluating scanpath models and underscore the importance of reporting design choices in GUI saliency research, with publicly available code and models to promote replicability.

Abstract

Modeling visual saliency in graphical user interfaces (GUIs) allows to understand how people perceive GUI designs and what elements attract their attention. One aspect that is often overlooked is the fact that computational models depend on a series of design parameters that are not straightforward to decide. We systematically analyze how different design parameters affect scanpath evaluation metrics using a state-of-the-art computational model (DeepGaze++). We particularly focus on three design parameters: input image size, inhibition-of-return decay, and masking radius. We show that even small variations of these design parameters have a noticeable impact on standard evaluation metrics such as DTW or Eyenalysis. These effects also occur in other scanpath models, such as UMSS and ScanGAN, and in other datasets such as MASSVIS. Taken together, our results put forward the impact of design decisions for predicting users' viewing behavior on GUIs.
Paper Structure (27 sections, 9 figures, 4 tables)

This paper contains 27 sections, 9 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. Dataset
  5. Design parameters
  6. Image size
  7. IOR decay
  8. Masking radius
  9. Evaluation metrics
  10. Dynamic Time Warping (DTW)
  11. Eyenalysis
  12. Determinism
  13. Laminarity
  14. Experiments
  15. Sensitivity to input image size
  16. ...and 12 more sections

Figures (9)

  • Figure 1: Impact of resizing to square or non-square image on different GUI types. The height is always fixed to 225 px. We consider widths of 128, 225, and 512 px. The best results are observed for widths of 225 px (resulting in a square aspect ratio).
  • Figure 2: Impact of resizing to different square image sizes on different GUI types. We consider sizes of 128, 225, and 512 px. The best results are usually observed for the 128 px cases.
  • Figure 3: Impact of different $\gamma$ values on different GUI types. Lower $\gamma$ means a high probability of revisiting fixation points. The best results are observed when $\gamma=0.1$.
  • Figure 4: Impact of different masking radius on different GUI types. Masking radii are relative to the input image size (e.g. 0.2 means 20% of the size). The best results are observed when the radius is set to 0.05.
  • Figure 5: Impact of different IOR mechanisms, using optimal parameters, on different GUI types. "'Baseline IOR" uses DeepGaze++ with the original IOR decay and the optimal parameters. "'Improved IOR" uses DeepGaze++ with our proposed IOR decay and the optimal parameters. The best results are usually observed with the baseline IOR.
  • ...and 4 more figures