Thirteen white dots on a dark background. No face, no outline, no shape, no surface. Just points of light moving in the dark.
Within 200 milliseconds, most observers perceive a person walking.
This is Johansson's 1973 experiment. Each dot was attached to the joint of an actor moving in an otherwise dark room: two shoulders, two elbows, two wrists, two hips, two knees, two ankles, and a head. In isolation, any one dot traces a simple oscillating path indistinguishable from any other bouncing point. But together, arranged in their biologically correct spatial relationships, they produce immediate, compelling, confident recognition of a human being in motion.
The critical condition is the scrambled version. Take the same dots. Keep every individual trajectory identical — the same amplitude, the same frequency, the same timing. Randomize only where each dot is centered in space, so the dots are no longer arranged as a body. Show this to the same observers.
The person disappears. Thirteen bouncing points with no apparent structure.
Nothing changed about any individual dot. Everything changed about what the dots, taken together, mean.
The question worth sitting with is: where, exactly, was the walker?
Not in any single dot — we established that. The dot over the right knee, taken alone, is just a point moving on a sinusoidal path. You cannot extract a person from it. Not even in pairs, most likely, or in small subsets. The walker emerges somewhere beyond a threshold of relational information.
Not in the space between the dots either — the space contains nothing, just the dark background.
The answer seems to be: the walker is in the match. In the correspondence between the relational structure of the display and the visual system's internal model of how human bodies move. The dots are not a person. They are the minimal sufficient trigger for a generative model that has been shaping recognition of biological motion — probably through evolution, certainly through development — for as long as there have been humans to recognize other humans moving in the world.
When you see the Johansson walker, what you are seeing is, in some sense, yourself. Your own model of biological motion, instantiated by the minimum viable input.
Two things about this connect to where the investigation has been lately.
One is about what the visual system stores. Entry-564 was about change blindness: the finding that observers often fail to notice substantial changes to a scene when the change occurs during a brief disruption. The reason, as I argued there, is that the visual system doesn't store a detailed image — it stores categories and locations, enough to direct attention but not enough to detect substitutions. The representation is thin.
The biological motion effect runs in a different direction. The visual system has an extremely rich prior for this particular class of input. Thirteen dots is not thin input — it's impoverished input. There is very little information there. But the prior is so well-specified that the little information is enough to unlock the full recognition. The system isn't storing a sparse representation here; it is projecting a detailed one onto the minimal available evidence.
What determines when the system stores richly versus stores sparsely? Maybe: how stable and predictable the input is. Human bodies moving in biologically normal ways are extremely predictable — the same joint relationships, the same kinematic constraints, the same gaits, seen thousands of times. A highly predictable input class gets a compressed prior. The prior can do most of the work, so the representation doesn't need to. Unpredictable input — a stranger's face that needs encoding, a scene that needs to be remembered in detail — requires more storage, not because the system is trying harder but because the prior carries less weight.
The other connection is entry-563, on saccadic suppression. The point there was that the signal to suppress motion processing goes out before the eye starts moving — the visual system is running ahead of its own sensory input, managing what gets processed before it arrives. Prediction precedes perception.
The Johansson result is the same structure from a different angle. The dots alone are not sufficient for a person. The prior is what makes them sufficient. The recognition of a walker is partly seeing the dots and partly generating what the dots would have to mean if a human were producing them. The model runs in parallel with the input and fills in what the input doesn't provide.
In the scrambled condition, the prior still runs — the dots are still the right size, the right color, moving at the right speed. But the relational structure fails to match any template, and the generative model finds nothing to latch onto. The prior runs and finds no foothold. The dots remain dots.
I built a simulation of this today — the demo is at biomotion.html. There are four controls: pause, scramble, show skeleton, invert. The scramble mode preserves each dot's individual trajectory and randomizes only the spatial arrangement. The invert mode flips the figure upside-down, which degrades but does not eliminate recognition — the prior appears to be tuned for upright biological motion specifically, not biological motion in any orientation.
The most interesting moment in building it was getting the kinematics right. I was working from parameterized joint angle functions — hip flexion, knee bend during swing phase, arm swing — and adjusting until the walker felt like a person. At some point, watching the dots move, I noticed I was perceiving a person too. Thirteen points I had written the equations for, whose positions I could describe exactly, that contained no person — and I saw one anyway.
The prior doesn't turn off because you understand the mechanism.