Adapting Color Difference for Design

CIELAB is commonly used in design as it provides a simple method for approximating color difference. However, these approximations model color perception under laboratory conditions, with correctly calibrated displays and carefully constrained viewing environments that are not reflective of complexity of viewing conditions encountered in the real world. In this paper, we present a data-driven engineering model for parametric color difference that extends CIELAB to be more broadly applicable to real-world conditions. Our model can be tuned to a desired range of viewers and conditions using a simple modeling procedure, while minimally increasing the complexity of the model. We demonstrate our approach empirically by modeling color differences for the web by leveraging crowdsourced participants. Introduction Color difference models are often used in design applications to predict how noticeably two colors will differ. These models serve several purposes, such as determining sets of colors that are subtly different (Fig. 1); however, they model perception under laboratory conditions, with correctly calibrated displays and constrained viewing environments. Given the rapid proliferation of visual content on the web and the increasing mobility of digital devices, visual media is becoming increasingly diverse, making factors that influence color difference perception, such as lighting conditions and display properties, highly variable in everyday viewing. Existing color difference models, while powerful descriptors of human vision, do not consider this variability, limiting their utility in design. CIELAB is commonly used in design scenarios as it offers a color difference formulation based on Euclidean distance (ΔE∗ ab). This metric is not as accurate as other appearance models, such as CIECAM02 [1], but its simplicity makes it practical for design. In this work, we present an approach to adapt CIELAB to model color difference perception for real-world viewing populations that preserves its simplicity. As the range of viewing factors in these populations is too complex to model each independently, we capture these factors in aggregate by sampling difference perception across target viewers and use these samples to derive scaling factors for CIELAB. The resulting model parameterizes CIELAB with respect to a given population and desired level of noticeable difference. Providing a parametric model tuned empirically to the target population exchanges the ability to make exacting claims about perceptual mechanism to instead create an engineering model that captures color difference in practice. This engineering model has several desirable properties for designers. It is parametric as it can be tuned to reflect a desired range of viewers and conditions. It is data-driven as it derives these parameters from observations under the target viewing conditions, yet practical as this data can About Tableau maps: www.tableausoftware.com/mapdata Sheet 1