Situation-Specific Models of Color Differentiation
David R. Flatla, Carl Gutwin · 2012 · ACM Transactions on Accessible Computing · doi:10.1145/2399193.2399197
Summary
This paper introduces Situation-Specific Modeling (SSM), a fundamentally different approach to helping users with color vision deficiency (CVD) differentiate colors in digital interfaces. Traditional recoloring tools rely on theoretical models of dichromatism—complete absence of one cone type—but this represents only a fraction of CVD cases. Most people with CVD have anomalous trichromacy (shifted cone sensitivities), and many others experience acquired deficiencies from cataracts, aging, diabetes, or environmental factors like tinted glasses or bright sunlight. Standard tools fail for these users because they model the wrong condition. Rather than simulating a theoretical CVD type, SSM empirically measures what colors a specific user can actually differentiate in their current environment. The approach builds individualized color differentiation (ICD) models through a short calibration procedure where users perform visual tasks, and the system records their just-noticeable difference thresholds. This captures all factors affecting color perception—genetic, acquired, and environmental—without requiring knowledge of the underlying cause. The paper presents four progressively refined ICD models. ICD-1 establishes the core approach using RGB color space, validated with 16 participants (half with CVD). ICD-2 switches to the perceptually uniform L*u*v* color space and uses discrimination ellipsoids rather than boxes, reducing calibration time from 32 minutes to about 2 minutes while improving accuracy. ICD-3 dramatically speeds up prediction time by using a single global ellipsoid that is translated and rotated rather than reconstructed for each color comparison. ICD-4 extends the binary differentiability prediction to a continuous degree-of-differentiability scale.
Key findings
The final ICD model achieves practical real-world performance: calibration takes approximately 2 minutes, and a 64-color image can be recolored in about 1 second. In user studies, ICD-2 achieved 78.7% overall accuracy—significantly better than ICD-1's 76.1%—while requiring 24 times less calibration time. ICD-3 made predictions 56 times faster than ICD-2 with no loss in accuracy, enabling near-real-time constraint optimization for recoloring algorithms. The models proved robust across environmental variations. Testing under different lighting conditions, background colors, and monitor calibrations showed that model accuracy degraded only moderately, meaning users don't need to recalibrate for minor environmental changes. The approach worked equally well for participants with and without congenital CVD, though CVD users required slightly larger safety margins in predictions. A key finding is that discrimination ellipsoids (ICD-2+) better match human color perception than the rectangular discrimination boxes of ICD-1, since human color confusion follows elliptical patterns along specific confusion lines in color space. The degree-of-differentiability function in ICD-4 enables smoother gradient recoloring and better optimization for continuous color scales, addressing limitations of binary differentiability predictions.
Relevance
This research has direct implications for building accessible data visualizations and interfaces. The 5% of the population with congenital CVD, plus those with acquired or situational color perception issues, regularly struggle with color-coded charts, maps, status indicators, and link differentiation. The ICD models provide a foundation for recoloring tools that actually work for real users in real environments. For practitioners, the key insight is that generic CVD simulations (like protanopia or deuteranopia filters) don't represent most affected users. A "one-size-fits-all" accessible color palette will inevitably fail for some users. The SSM approach suggests that personalized color adaptation—while requiring brief calibration—can achieve far better results than generic solutions. The software implementing these models was made publicly available (hci.usask.ca/ICD/), providing a practical tool for researchers and developers. The broader SSM methodology also extends beyond color vision: the authors suggest it could apply to any accessibility domain where user capabilities can be empirically measured and modeled, including hearing difficulties and low visual acuity. This positions situation-specific modeling as a general framework for personalized assistive technology.
Tags: color vision deficiency · color blindness · recoloring · visualization · personalization · low vision · assistive technology