Only Right Way Up
You own a piece of hardware that does faces and only faces. It has an orientation. This page turns its input over and measures the size of the hole that leaves.
Two faces, side by side, for about a second. Sometimes they are the same face twice. Sometimes one of them has been changed. You say which.
Half the time the change is spacing: identical parts, moved. Eyes a little further apart, a mouth a little lower. Half the time it is a feature: identical positions, different parts. A heavier brow, a thinner lip, a narrower eye.
And half of all of it is upside down. That is the whole experiment. Yin published it in 1969 and it is still the cleanest evidence that face perception is not just object perception aimed at a face: turn a house over and you get worse, turn a face over and you fall off a cliff.
- 0️⃣ Three practice pairs with the answer shown after each. Nothing is scored.
- 1️⃣ Forty eight pairs, shuffled, twelve in each of four cells: spacing and feature, upright and inverted.
- 2️⃣ One question every time: same or different. Keys S and D work too.
- 3️⃣ Exactly half of every cell is genuinely identical, so answering different at everything scores dead chance.
- 4️⃣ The headline is how much of your own upright ability survived the flip, per kind of change.
- Nothing is scored as an absolute. Every headline is a ratio against your own upright cell, which shares the faces, the exposure, the screen and your fingers. How good you happen to be at faces is in the denominator and it cancels.
- Same and different trials are equal in number in every cell, and the score is balanced accuracy, hits and correct rejections averaged. Liking one button cannot buy you anything. Your false alarm rate is printed anyway.
- A new random face is generated for every single trial. No face appears twice, so there is nothing to learn and nothing to memorise.
- Inverted means a 180 degree rotation of the exact same drawing. Not a different face, not a different size, not a redraw. Every pixel that was there is still there.
- The known trap in this literature is that the two kinds of change must be about equally hard upright, or a "selective" loss is just a ceiling. Both upright numbers are printed side by side, and if they are more than fifteen points apart this page says so and refuses to headline the comparison.