How cats see color

Guenther and Zrenner (1993) recorded from ganglion cells in cat retinas. In daylight conditions every cell received input from a long-wave cone system peaking at 550 nm, but only 15% received input from a short-wave system peaking at 450 nm. In the dark, 92% of the cells were driven by rods, with a peak at 501 nm.

Most electrophysiological studies found two kinds of cone in cats, near 560 and 460 nm, and a few reported a third receptor at 500–520 nm. Clark and Clark (2016) tested two cats with single-wavelength light mixed with white of varying brightness. The cats told the colored light from white between 456 and 497 nm and between 510 and 524 nm, but failed at 505 nm. A color that cannot be told from white is called a neutral point, and a neutral point is the mark of a dichromat. Theirs, near 505 nm, is almost the same as that of a person with deuteranopia.

The simulator is built from the cat's own cone peaks (450 and 550 nm). For each color it estimates how strongly it would excite each of the two cones, using the pigment template of Govardovskii and colleagues (2000) and daylight (CIE D65), and shows a color that excites them in the same way. The cone signals weigh the light by its energy; a cone counts photons, and weighting by photon count instead would shift some shown colors slightly, by up to about 16 of 255 in one color channel by our check. The shown colors lie on a yellow–blue axis plus lightness, the same plane that Viénot, Brettel and Mollon (1999) use for people with two cones, so colors that give the cat the same pair of cone signals come out identical. In this model the cat's neutral point falls near 495 nm, about 10 nm short of the 505 nm that Clark and Clark (2016) measured. It is an approximation: the eye's lens is left out, each color's spectrum is a guess from its screen color, and the picture is blurred to the animal's visual acuity, as explained below. Everything runs on your device: the camera picture, photos and videos are not uploaded.

Pasternak and Horn (1991) measured the finest striped grating that cats could detect while they held their gaze on a spot of light. In the center of the eye, the area centralis, the limit was about 3.5 cycles per degree, and it fell further from the center. The simulator blurs the picture to 3.5 cycles per degree. We read only the summary (abstract) of the paper, so the light level of the test is not known to us.

How the blur is made: the simulator follows AcuityView, the method of Caves and Johnsen (2018). Fine detail is faded along a bell-shaped curve that leaves about 3% of the contrast at the animal's acuity limit, and coarse shapes and colors stay. A picture does not say how wide the scene was, so the simulator assumes that the long side of the picture covers about 70°, the view of a typical phone's main camera (the iPhone 16 main camera is 26 mm in 35 mm terms, about 69° across). If you zoomed in or cropped, the scene was narrower and the animal would see it blurrier than shown. The blur stands for the sharpest part of the eye only; the rest of the animal's view is not modelled.

Better than we can, but not as a bright, colorful world. Kang and colleagues (2009) tested cats and people in the same task: spotting a faint striped patch to one side of where they looked, at light levels spanning eight orders of magnitude, from a millionth of a candela per square metre up to about 60 cd/m². Cats detected large, coarse shapes in light about 5 to 6 times dimmer than people needed (0.74 log units). Most of that advantage came from the optics of the eye, its large pupil and short focal length. The tapetum, the reflective layer behind the retina, added about 29% by the authors' estimate from older measurements, not by a test of their own. Fine detail was no better: cats and people both lost stripes of 4 cycles per degree at about the same dim level, around 0.004 cd/m².

Colors fade in dim light for people too. In a test run on horses and people, most people still told colors apart at 0.02 cd/m², a level the authors compare to moonlight, and failed at 0.007 (Roth and colleagues, 2008).