What this tests: Scrutinizer's chromatic pooling model predicts that color information
decays faster than luminance in peripheral vision, with red-green (RG) channels collapsing ~5×
faster than blue-yellow (BY). These predictions derive from castleCSF parameters (Bowers et al. 2025)
applied to Oklab perceptual color space. We validate by rendering colored dot arrays through
Scrutinizer's filter, measuring chroma retention at each eccentricity ring, and comparing
the RG/BY decay ratio against published psychophysical data from Mullen & Kingdom (2002),
Bowers (2025), and Hansen et al. (2009).
Predicted chroma retention per color at each eccentricity ring.
Red and green decay fastest (RG channel dominates), blue and yellow slower (BY channel).
Green tracks the RG curve, not BY — a non-obvious prediction from its Oklab a-axis projection.
Per-Channel Retention: RG vs BY at 24px
Isolating the two chromatic channels: RG collapses ~5× faster than BY.
Dashed lines show Mullen & Kingdom (2002) published sensitivity. Open circles show Bowers (2025).
The model tracks published data within the 20% tolerance at matched eccentricities.
Hansen 2009: Naming Accuracy vs Model Retention
Hansen et al. (2009) measured how accurately people name colors at different eccentricities.
Solid lines show our model's chroma retention; dashed lines show Hansen's naming accuracy.
If chroma retention predicts naming ability, these curves should correlate (Tier 3 target: r > 0.8).
Measured Retention (filtered) vs Model
Screenshot measurements overlaid on model predictions. Measured values are compressed
toward the low end because Mode 0's spatial blur applies to both conditions, reducing the
dynamic range before chromatic pooling acts. The relative ordering should still match.
Tier 1: Must Pass
Observation: Chroma retention must monotonically decrease with eccentricity for all colors,
and BY retention must exceed RG retention at the outermost ring. These are fundamental predictions
of the chromatic pooling model — if these fail, the model is wrong.
Observation: The RG/BY decay ratio should match published psychophysical data within 20%.
Green should track the RG curve (Oklab a-axis), not the BY curve — a prediction that distinguishes
our Oklab-based model from naive hue-based approaches.
FAILBY/RG channel ratio vs Bowers at ~15°: model=2.15 (yv/rg retention) vs Bowers=2.72 (21% off, threshold=20%)
PASSGreen closer to red than blue: green-red gap=8.8pp, green-blue gap=21.6pp (threshold: <15pp and closer to red)
FAILRendered matches model within 15%: 2/10 (20%)
Tier 3: Stretch
Observation: Does our chroma retention curve predict real-world color perception tasks?
Hansen et al. (2009) measured color naming accuracy across eccentricity. If our model captures the
underlying signal, the correlation should be strong (r > 0.8).
PASSred model retention correlates with Hansen naming accuracy: r=1.000 (threshold: r>0.8)
PASSblue model retention correlates with Hansen naming accuracy: r=1.000 (threshold: r>0.8)