Solving an Invisible Problem: Designing for Color-Blindness in Games

• Speaker: Douglas Pennant (spelling as transcribed; the surname may be "Pennant" or similar), associate development manager at Creative Assembly, himself a deuteranope. Earlier QA work on Alien: Isolation and Kinect Star Wars; Halo Wars 2.
• Year: not stated. Most likely GDC 2019, inferred from "Fortnite, as of November, 125 million players", a talk at the Game Accessibility Conference that week about the Xbox Adaptive Controller, and Halo Wars 2 as his team's last release.
• Video: https://www.youtube.com/watch?v=KbFs9ghIIEI

What the talk teaches

1. The problem is invisible to whoever has it. A colour-blind tester cannot report a cue he never sees (the Working Joes' eyes in Alien: Isolation changed colour when hostile; he learned it a month or two before ship), and people with normal colour vision do not know it is broken. Someone has to start the conversation and test on purpose.
2. Cover the three dichromacies (protanopia, deuteranopia, tritanopia): they are the extreme ends of the milder anomalous kinds, so supporting them covers those too. About 1 in 12 men and 1 in 200 women are colour blind; deuteranopia is about three quarters of colour-blind men, protanopia about a quarter, tritanopia about 1 in 10,000 and equal across sexes.
3. Where games go wrong: team colours, units that blend into the ground, colour-matching puzzles, health shown as green-yellow-red, crowded minimaps, racing lines that turn red to brake, glowing eyes as the only sign of aggression.
4. Never colour alone (WCAG 1.4.1): Hue adds symbols, FTL crosshatches rooms without oxygen, Mass Effect keeps its two dialogue choices in fixed places.
5. Blue and orange for pairs that must never be confused. In his words: "If you have two things and you can't have them get mixed up, blue and orange, every time." Not for total colour blindness, unless the brightness differs too.
6. Semantic colour presets. Pick colours for friendly, enemy and faction A once, from a colour-safe palette; artists and designers tag by meaning; a preset that fails a test is fixed in one place. Test the presets in context, with simulators (Color Oracle, Sim Daltonism, Coblis, Unity and Unreal packages) and with colour-blind people.
7. Let players pick the colours (AudioSurf, Overwatch's enemy and friendly colours, Battlefield 1's hue picker). Asked whether he prefers a whole-screen correction filter: no, change the UI; "I don't need the entire screen to change color."
8. Make the options findable (Fortnite's Accessibility tab). Its tritanopia mode is used by many players who are not tritanopes, because it reads more clearly.

What the kit took

• shot -ColorVision protan,deutan,tritan (ci/godot/tools/shot.gd, ci/godot/lib/Measure.ps1, ci/godot/hyperdot.ps1): every screenshot is also saved as <name>.protan.png, .deutan.png, .tritan.png, listed in shots[].colorVision (docs/json.md). The model is Machado, Oliveira and Fernandes (2009), severity 1.0, applied in linear RGB: sRGB decode, matrix, clamp, encode. The matrices are the authors' published table; their page (inf.ufrgs.br/~oliveira) could not be reached from here, so they were checked against three implementations that agree to six decimals: colorspacious (cvd.py, which says it downloaded them from that page), colour-science (blindness/datasets/machado2010.py) and DaltonLens-Python (simulate.py).
• tests/Test-ColorVision.ps1 with tests/colorvision-fixture (in the measure job of self-test.yml).
• ui-accessibility: colours defined once by meaning, blue and orange for must-not-confuse pairs, player-chosen marker colours, recolour markers and UI rather than filter the screen, and the -ColorVision check. look: critique the simulated pictures.

Measured

• Fixture swatches, CIE76 ΔE between brick red #C84610 and leaf green #328200: 94.4 seen normally, 12.0 as a protanope, 11.4 as a deuteranope. Blue #377EB8 and orange #FF7F00: 121 normally, 99, 115 and 98 in the three simulations. Every simulated swatch is within 1/255 of the matrix computed in PowerShell; grey stays grey.
• The red and green the triage first proposed (ColorBrewer #E41A1C and #4DAF4A) stay 32 (protan) and 21 (deutan) apart, because they differ in lightness: brightness alone still separates them, which is why the fixture uses a pair of similar lightness.
• Applying the matrix to sRGB values directly (the common mistake) is off by up to 26/255 and understates the collapse: red and green come out 20 apart for a protanope instead of 12. The self-test fails on that version (13 checks).
• Speed at 1920x1080 under Xvfb with software OpenGL (llvmpipe), Godot 4.7.2: a per-pixel GDScript loop takes 1.3 s per kind; the shader, one SubViewport per kind, takes about 250 ms for all three including the read-back, and saving the three PNGs about 300 ms. A GPU makes the shader part faster still.

What it did not take, and why

• Achromatopsia (about 1 in 30,000) and the milder anomalous kinds: the talk covers the dichromacies on purpose, and severity 1.0 is the worst case of each.
• DaltonLens notes that Machado's model is less accurate for tritanopia than Brettel (1997); the tritan picture is a guide, not a diagnosis. A colour-blind tester remains the real check, as the talk insists.
• A colour-blind mode or colour picker in the C# example: the rule is in the skill; the kit's part is letting an agent see the screens.