Guidelines
Suggest edit

Add a low-center-bias subset panel

For grouped-result comparisons of fixation-prediction models on center-biased datasets, use a low-center-bias subset panel in a multi-view benchmark to improve fidelity and address over-crediting center-focused models for analysts.

  • purpose:refine
  • basis:empirical
  • task:compare
  • scope:grouped-result
  • structure:multi-view
  • quality:fidelity:use
  • lever:layout-structure
  • communication:credibility

advice

Low-center-bias subset view

Add a separate low-center-bias subset panel when the full fixation dataset is strongly centered. For example, compute a center-bias ratio from pooled fixations, select images with CBR below the paper’s threshold at the 40% radius level, and show the subset ranking beside the full-dataset ranking.

reason

Why the subset panel matters

Strong central clustering can hide which methods really predict off-center gaze. A low-center-bias subset surfaces the harder cases that are more informative for comparison.

Mechanism: The subset panel reduces the influence of central fixation density, so readers can judge how methods behave on off-center fixations rather than on the dataset’s viewing bias.

Evidence: The paper introduces a center-bias ratio, uses a 40% radius criterion with a threshold of 0.7 to select less center-biased images, and shows that this second-order analysis helps differentiate model behavior on off-center fixations while the centered Gaussian performs poorly there (Borji et al., 2013).

Notes: The paper reports that only a small fraction of images passed this filter in the tested still-image datasets.

context

Use when the full benchmark is center-heavy

  • User Goal: Evaluate non-trivial fixation prediction rather than center preference.
  • Task: Compare models on natural-scene eye-tracking datasets.
  • Data: Pooled human fixations with strong center bias in the full dataset.
  • Chart Setting: Multi-view benchmark figure with room for a subset result beside the full result.
  • Audience: Analysts and reviewers checking fairness of model comparisons.
  • Success Criterion: Off-center prediction differences become visible.

exceptions

Do not use it when no valid subset exists

Break it when: No images pass the center-bias filter. Why: The paper reports a dataset where none of the images met the threshold, so a low-center-bias subset view could not be formed.

costs

What this costs

Sacrifice: The subset panel uses fewer observations than the full dataset. Risk: Scores may become noisier because the sample is much smaller. Mitigation: Show the subset size directly in the panel so readers can judge stability.

mistakes

Common subset-panel failure

Mistake: Replace the full-dataset result with the low-center-bias subset. Why it fails: The paper uses the subset as a second view, not as a full replacement for the original benchmark.

check

How to test whether the subset panel is needed

Failure Sign: The full-dataset ranking stays strong for a centered baseline or mostly reflects central fixation density. Quick Check: Compute the center-bias ratio at the 40% radius level and count how many images fall below 0.7. Stronger Test: Compare model rankings between the full dataset and the low-center-bias subset; if the order changes, keep both views.

fix

What to change

  • Pool fixations across viewers for each image without Gaussian smoothing before computing center bias.
  • Compute the center-bias ratio across concentric central circles and apply the 40% radius threshold used in the paper.
  • Add a separate subset panel with its own sample size.
  • If no images pass the threshold, omit the subset panel rather than forcing a tiny or invalid subset.

References

Borji, A., Sihite, D. N., & Itti, L. (2013). Quantitative Analysis of Human-Model Agreement in Visual Saliency Modeling: A Comparative Study. IEEE Transactions on Image Processing, 22(1), 55–69. https://doi.org/10.1109/TIP.2012.2210727