To better understand how local motion detectors merge their responses so as to permit the global determination of objects' movements in the visual field, direction discrimination of performance was measured using a flexible class of moving dots--two sets of dots translating sinusoidally 90 deg out of phase along orthogonal axes. When dots' velocities are combined, a global motion along a circular trajectory emerges, clockwise or counter-clockwise depending on the sign of the phase lag. However, the results of the present experiments indicate that dot patterns are segregated into distinct, but interacting, streams when each dot motion can be accurately determined. In contrast, perceptual coherence of the global motion occurs when each local motion signal is "blurred" by a "motion noise". Direction discrimination performance then increases regularly with both noise amplitude and noise frequency, i.e., noise speed. Performance also increases when relative motion between dots is added. Testing different dot configurations indicates that performance is better for spatial arrangements that display structural properties (a square shape), as compared to overlapping random distributions. Interestingly, when the delay between stimulus onset and motion onset increases up to 300 msec, performance improves when dot patterns convey come form of structural organization but not when the dots are distributed at random. Relations of these results to existing models of motion integration are considered.