Discrete windows of feature integration

A car drives through the night. It is hard to estimate its color from single photoreceptor activity because photoreceptor activation is short and noisy. An efficient way to estimate the color is to average photoreceptor activities along the car’s motion trajectory. Here, we show that humans indeed integrate visual features along trajectories for substantial times. We presented a vernier followed by subsequent pairs of flanking lines, creating the impression of two expanding motion streams. When a flanking line is offset, its offset integrates with the central vernier offset. This integration is mandatory for about 450 milliseconds, that is, observers cannot report the two offsets independently. Integration is even mandatory when observers make an eye movement during this period. Integration is precise, that is, vernier offsets from the two streams never integrate—just as colors of two cars are not mixed. However, when vernier offsets are presented in subsequent integration windows, observers can report both offsets independently. For example, the central vernier integrates mandatorily with the flank offset at 330 milliseconds but not with one at 490 milliseconds. Surprisingly, the flank offsets at 330 milliseconds and 490 milliseconds do not integrate although they are in close temporal proximity. We propose that integration comes with perceptual quants.