Comprehensive characterization of human color discrimination thresholds

This study overcomes the historical challenge of comprehensively characterizing human color discrimination thresholds by combining an adaptive trial-placement procedure with a novel semi-parametric Wishart Process Psychophysical Model (WPPM), enabling efficient mapping of discrimination performance across the isoluminant plane with high accuracy and broad applicability to other perceptual domains.

The Gist

Imagine you are trying to map the entire surface of a strange, invisible island. This island represents human color vision. Your goal is to find the exact spots on this island where a person can just barely tell the difference between two colors.

For a long time, scientists thought mapping this whole island was impossible. Why? Because the island is four-dimensional (it involves two colors for the "base" and two colors for the "difference"), and checking every single spot would take a lifetime of staring at screens. It's like trying to find every single grain of sand on a beach by picking them up one by one.

This paper describes a brilliant new way to map that entire island in just a few hours. Here is how they did it, explained simply:

1. The Problem: The "Dimensional Curse"

Think of color discrimination like a game of "Spot the Difference."

• Old Way: Scientists used to pick one specific color (like a specific shade of blue) and then test tiny changes around it. To map the whole island, they'd have to pick a new starting color, test again, pick another, and test again. It was slow, tedious, and they could only map tiny patches.
• The Challenge: Because there are so many possible colors, the number of tests needed to cover everything grows exponentially. It's the "curse of dimensionality."

2. The Solution: A Smart GPS and a Smooth Blanket

The researchers used a two-step strategy to solve this:

Step A: The Smart GPS (Adaptive Trial Placement)
Instead of walking randomly or checking every spot, they used a computer algorithm (called AEPsych) that acts like a super-smart GPS.

• Imagine you are looking for a lost coin on a beach. A normal person might walk in a straight line. This GPS, however, looks at where you just looked. If you found a coin nearby, it says, "Hey, let's look a little closer to that spot!" If you found nothing, it says, "Let's try a different area."
• The computer watched the participants' eyes and brains in real-time. If a participant was struggling to tell two colors apart, the computer made the next test slightly easier. If they were breezing through, it made it harder.
• Result: In just 6,000 trials (which took about 12 hours of work per person), they gathered the most important data points needed to understand the whole island, skipping the boring parts where the answer was obvious.

Step B: The Smooth Blanket (The WPPM Model)
Now they had a bunch of scattered data points (like dots on a map). They needed to connect the dots to see the whole picture.

• They invented a new mathematical model called the Wishart Process Psychophysical Model (WPPM).
• Think of this model as a smooth, stretchy blanket that they draped over the scattered dots.
• The Key Insight: The researchers assumed that human vision doesn't have "jumps." If you can barely tell the difference between a specific red and a slightly redder red, you should be able to tell the difference between a very similar red and a slightly redder red too. The "noise" in our brains changes smoothly, not randomly.
• The "smooth blanket" math forces the model to fill in the gaps between the dots in a way that makes physical sense. It predicts what the answer would be for any pair of colors, even ones they never actually tested.

3. The Results: A Complete Map

By combining the Smart GPS (to get the data) and the Smooth Blanket (to fill in the gaps), they created a complete, 4D map of human color discrimination.

• What they found: They confirmed that we are best at telling colors apart near "gray" (the center of the island). As colors get more saturated (brighter reds, blues, greens), it gets harder to tell them apart.
• The Shape of the Island: They found that the "difficulty" isn't a perfect circle. It's more like an oval. Depending on which direction you look (e.g., toward green vs. toward purple), the difficulty changes.
• Validation: To make sure their "Smooth Blanket" wasn't just making things up, they set aside 6,000 new tests they didn't use for the map. They compared their predictions against these new tests, and the predictions were almost perfect.

4. Why Does This Matter?

This isn't just about colors; it's about how we see the world.

• Better Screens: This data helps engineers design phone and TV screens that use the exact right amount of color to look perfect to the human eye, saving energy and money.
• Eye Health: Doctors can use this map to detect eye diseases earlier. If a patient's "map" looks different from the healthy average, it could be an early sign of trouble.
• AI and Vision: It gives computer scientists a "gold standard" to test if their AI vision systems are seeing the world like humans do.

The Takeaway

The authors took a problem that was considered "hopelessly difficult" (like trying to count every grain of sand on a beach) and solved it by using a smart guide to find the interesting spots and a mathematical blanket to connect them. They didn't just map a few patches; they mapped the entire territory of human color vision, efficiently and accurately.

Technical Summary

1. Problem Statement

The central challenge addressed in this study is the "psychophysical curse of dimensionality" regarding the characterization of human color discrimination thresholds.

• The Complexity: While color discrimination is often studied in single dimensions (e.g., contrast), a full characterization requires mapping a four-dimensional psychometric field. This field defines the probability of a perceptual response as a joint function of a reference stimulus and a comparison stimulus, both varying across two color dimensions (in an isoluminant plane).
• The Limitation: Traditional methods, such as the Method of Constant Stimuli (MOCS), require an exponential number of trials to tile this high-dimensional space, making a comprehensive map practically intractable.
• The Gap: Existing adaptive methods often rely on parametric assumptions about the underlying noise structure that may not be known a priori, potentially biasing the data collection.

2. Methodology

The authors propose a novel framework combining non-parametric adaptive sampling with a semi-parametric Bayesian model to efficiently map the psychometric field.

A. Experimental Design

• Task: A 3-Alternative Forced Choice (3AFC) oddity task. Participants viewed three "blobby" 3D objects (two identical references, one different comparison) and identified the odd one out.
• Stimuli: Naturalistic 3D objects rendered in Unity, constrained to the isoluminant plane of the display (constant luminance, varying chromaticity).
• Participants: 8 individuals with normal color vision.
• Trial Count: Each participant completed ~6,000 trials for model fitting and an additional 6,000 held-out trials for validation.

B. Data Collection: Non-Parametric Adaptive Sampling

• Tool: Used AEPsych, an open-source package for adaptive psychophysics.
• Mechanism: Employed a Gaussian Process (GP) model with a Radial Basis Function (RBF) kernel to assume smooth variation in performance across the stimulus space.
• Strategy:
  • The first 900 trials used quasi-random Sobol' sampling for initialization.
  • The remaining 5,100 trials were adaptively selected using the Expected Absolute Volume Change (EAVC) acquisition function to target the 66.7% correct threshold level.
  • Crucial Feature: This approach is "non-parametric" relative to the final model, meaning it did not assume the specific form of the internal noise (covariance structure) during data collection, preventing bias.

C. The Core Model: Wishart Process Psychophysical Model (WPPM)

The WPPM is a Bayesian probabilistic model fitted post hoc to the adaptively collected data.

• Observer Model: Assumes internal representations of stimuli are noisy multivariate Gaussians. The observer identifies the "odd" stimulus by comparing pairwise Mahalanobis distances between internal representations, accounting for the noise structure.
• Smoothness Prior: The core innovation is modeling the covariance matrix field (representing internal noise) as a finite-basis Wishart Process.
  • The covariance matrix $\Sigma(x)$ at any stimulus location $x$ is derived from a weighted sum of 2D Chebyshev polynomial basis functions.
  • A Wishart process prior is placed over the weights of these basis functions. This enforces the assumption that internal noise varies smoothly across the stimulus space (small changes in stimulus $\rightarrow$ small changes in noise).
  • Hyperparameters ($\epsilon$ and $\gamma$) control the strength of this smoothness regularization.
• Fitting: The model parameters (weights) are estimated via Maximum A Posteriori (MAP) using gradient-based optimization (stochastic gradient descent) on the log-posterior density, approximated via differentiable Monte Carlo simulations.

3. Key Contributions

1. First Comprehensive Map: The study provides the first high-resolution, comprehensive characterization of human color discrimination thresholds across the entire isoluminant plane for multiple observers.
2. Novel Modeling Framework (WPPM): Introduces a semi-parametric model that leverages the smoothness of internal noise to infer a continuous 4D psychometric field from sparse data.
3. Efficiency: Demonstrates that ~6,000 trials per participant are sufficient to map the entire field, a feat previously considered "hopelessly difficult" (referencing Schrödinger, 1920).
4. Validation Strategy: Validates the model against 25 independent probe psychometric functions (measured via MOCS) held out from the model fitting, ensuring the model generalizes beyond the training data.

4. Results

• Model Accuracy:
  • The WPPM predictions showed high agreement with the independent validation thresholds (mean correlation coefficient = 0.84; slopes near 1.0).
  • 95% bootstrap confidence intervals of the WPPM predictions overlapped with validation thresholds in 22–25 out of 25 conditions for most participants.
  • Simulations using a ground-truth WPPM confirmed the model's ability to recover the underlying field with high fidelity.
• Observed Regularities:
  • Thresholds: Lowest near the achromatic (gray) point and increasing with distance from it.
  • Orientation: The major axes of the elliptical threshold contours tend to be radially oriented toward the achromatic point.
  • Individual Variability: While general patterns were consistent, significant individual differences were observed, particularly in the orientation of ellipses in high-saturation regions.
• Comparison with Literature:
  • The results qualitatively align with classic MacAdam (1942) ellipses and recent studies (Danilova & Mollon, 2025; Krauskopf & Gegenfurtner, 1992).
  • Metric Validation: The measured thresholds showed poor agreement with the standard CIELab $\Delta E_{76}$ metric but much better alignment with the more modern $\Delta E_{94}$ and $\Delta E_{00}$ metrics, suggesting limitations in older color difference formulas for predicting threshold-level discrimination.

5. Significance and Implications

• Foundational Dataset: Provides a benchmark dataset for testing computational and neural models of color vision, moving beyond simplified 1D or sparse 2D characterizations.
• Generalizability: The approach (adaptive sampling + smoothness-constrained modeling) is not limited to color. It can be applied to any perceptual domain where internal noise varies smoothly across stimulus space (e.g., motion, auditory speed, numerosity).
• Supra-threshold Metrics: The data supports the hypothesis that color space can be modeled as a Riemannian manifold (locally Euclidean but globally curved), where supra-threshold differences are the accumulation of local threshold differences (geodesics).
• Clinical and Industrial Applications: The method offers a scalable, data-efficient tool for quantifying eye diseases, optimizing display technologies, and refining perceptual metrics used in cognitive science and computer vision.

In conclusion, this paper successfully overcomes the dimensionality barrier in psychophysics by combining adaptive data collection with a smoothness-prior-based generative model, yielding a precise, comprehensive map of human color discrimination capabilities.