Three cars, obviously; their size increases from bottom left to top right. Actually, I have only one of these cars, there are simply two identical copies of the car photo inserted at different places in the lattice bridge.
Hover with the mouse over the image, or tap it. Then the background is hidden, and the illusion becomes weaker.
This is closely related to Shepard’s “Subterranean Terror” and a variant of the “Ponzo Illusion” (see figure below right), where image spaces are inappropriately mixed. The background, the lattice bridge, is distorted according to the rules of perspective: Everything farther away in real space is smaller in image space. In this image space I inserted 3 identical car images. Because they are not perspectively distorted, but our brain assumes they are(because normally everthing is), it is quite a correct perceptual outcome that the cars grow in size from bottom left to top right. This is essentially what Richard Gregory (1963, 1988) called “inappropriate constancy” – only that I find it “appropriate” here.
In normal viewing this “size constancy”, namely the perceptual correction of retinal image size (e.g. vertically across 1000 ganglion cells) by (assumed) distance, is always in operation. It prevents, for instance, the perception that someone walking away from us becomes a dwarf.
There are a number of different hypotheses, reviewed in Yildiz et al. (2021). These authors concluded that a slightly modernised (“Bayesian”) version of Gregory’s explanation best covers all experimental findings.