Read the score before trusting it
Glance estimates early visual attention. Each score answers a narrower question than “does this design work?”
What goes into a heatmap
Let pᵢ be normalized, nonnegative predicted fixation density at pixel i, with Σpᵢ = 1. An area of interest (AOI) is a region you draw. Attention share describes predicted map mass, not a percentage of viewers.
Scores for reviewing a design
| Output | Formula or method | How to read it |
|---|---|---|
| AOI share | 100 × Σ(i ∈ AOI) pᵢ | Predicted attention mass inside an area. Overlapping areas can double count. |
| Clarity | 100 × [1 − H(p)/ln(4096)] H(p) = −Σpᵢ ln(pᵢ) | Downsample to 64 × 64 and renormalize first; 0 ln 0 = 0. Concentrated maps score higher. This custom 0–100 index is not validated comprehension. |
| First glance | Rank each AOI by the first estimated fixation that enters it. | Unvisited areas have no rank. Fallback: seven peaks, Gaussian inhibition of return and center prior. Approximation; width-based suppression is not calibrated visual angle. |
| Fog | Dim below the specified “25th percentile mass” threshold. | Low predicted attention, not proof of being ignored. TODO-verify: whether implementation uses a pixel quantile or cumulative mass threshold. |
| Agreement | Uᵢ = stdₘ(pₘᵢ) scalar = mean pairwise CC | Ensemble maps aligned to the same grid. High agreement can accompany shared errors. Constant maps make CC undefined. |
| Compare A/B | Δshare = share(B) − share(A) Δclarity = clarity(B) − clarity(A) | Share deltas in percentage points; report heatmap CC separately. Prediction differences are not causal conversion lift. |
| Video | UNISAL temporal mode; sample at 5 frames/s. | Per-frame heatmaps and AOI shares over time. Target behavior; export and timing verification pending. |
Metrics for comparing predictions with people
Let S be a predicted saliency map, G a human fixation density and F the set of recorded fixation locations.
| Metric | Formula | Meaning and limit |
|---|---|---|
| AUC | ∫₀¹ TPR(FPR) dFPR | How well map scores separate fixated locations from negatives; 0.5 is chance under the chosen protocol. Negative sampling and center bias matter. |
| CC | cov(S,G) / [σ(S) σ(G)] | Pearson correlation of map values, −1 to 1; larger is more similar. Undefined for constant maps. |
| NSS | mean(i ∈ F) [(Sᵢ − μ(S))/σ(S)] | Average standardized map value at human fixations; higher is better. It is not a percentage. Undefined when σ(S) = 0. |
Source: saliency metric review. Read the benchmark evaluation protocol before comparing results.
Reference benchmark, not a Glance result
| Reference | AUC | CC | NSS |
|---|---|---|---|
| DeepGaze IIE | 0.8829 | 0.8242 | 2.5265 |
| Human gold standard | 0.9341 | 0.9828 | Not supplied |
MIT/Tübingen benchmark · DeepGaze IIE paper. These values do not establish performance of SUM, TranSalNet, UNISAL or an exported ONNX model in Glance.
Commercial “92–96% accuracy” figures may normalize a metric against the human ceiling. For example, 83% / 90% ≈ 92%. That ratio does not mean “correct 92% of the time.” Request the raw metric, denominator, dataset and validation procedure for each claim.
Where prediction stops
- Early free viewing, roughly the first 3 seconds, differs from reading or goal-directed tasks.
- Training data may not represent your audience, image type or viewing context.
- Map concentration and model agreement do not measure understanding, preference, memory or emotion.
- Model export, resizing and preprocessing can change results. Export parity remains to be verified.
- Use real participants for decisions requiring actual behavior, accessibility or conversion evidence.
Questions before you start
Is Glance 94% accurate?
No such Glance result is established. Reference benchmark values describe a named research model and protocol, not this application.
Does higher clarity mean a better design?
No. Clarity measures concentration in the predicted map. A very concentrated map can overlook useful information.