Fangfang Hong¹ (), Ruby Bouhassira¹, Craig Sanders², Michael Shvartsman³, Alex Williams⁴, Phillip Guan², David Brainard¹; ¹Department of Psychology, University of Pennsylvania, ²Reality Lab Research, Meta, ³FAIR, Meta, ⁴Center for Neural Science, New York University

Discrimination thresholds are foundational for understanding the limits and mechanisms of color vision. Fully characterizing these thresholds has been intractable because, with standard methods, trial number grows exponentially with stimulus dimensionality. We overcome this challenge using adaptive trial placement (AEPsych, Owens et al. 2021) and a semi-parametric Wishart Process model (WPM). The WPM leverages smoothness in performance variation and enables interpolation of thresholds across the full color gamut. As proof of concept, we measured color discrimination across the DKL isoluminant plane using a three-alternative forced-choice task. On each trial, participants (N = 5) identified the odd-colored stimulus. Nine reference stimuli were sampled, with comparison stimuli chosen around each reference by AEPsych (3,240 trials/participant). We also ran blocks of validation trials using the method of constant stimuli (MOCS) for both tested and untested reference stimuli (2,880 trials/participant). We fit the WPM to infer the covariance matrices of a multivariate Gaussian internal noise field. From the field, we predict percent correct discrimination for any reference and comparison. Thresholds from validation trials were highly correlated with WPM predictions, but were systematically lower (mean correlation = 0.878, mean ratio = 1.374). We think the bias is due to reduction in uncertainty for the MOCS trials. The data reveal consistent patterns across participants: (1) thresholds are smallest at the achromatic reference; (2) thresholds increase with reference distance from achromatic; (3) threshold ellipses are oriented toward the achromatic reference; (4) thresholds deviate substantially from predictions based on CIELAB. In sum, our approach enables a complete characterization of color thresholds across the isoluminant plane. We expect it will generalize to ellipsoids in full 3D color space, and the validation bias will diminish when MOCS trials are run fully interleaved. Moreover, our method induces a Riemannian geometry on color space that provides a candidate perceptual distance metric.

Acknowledgements: Meta