Human perception is largely non-retinotopic. A bike passes by; we perceive the reflectors on its wheels rotating on a circular orbit. However, the “real” motion is truly different, namely cycloidal. The real, retinotopic motion is invisible because the bike motion is subtracted from the reflector trajectory. Thus, the bike creates a non-retinotopic reference frame within which the reflector motion is perceived. Almost nothing is known about how this feat is achieved. Here, we acquired high-resolution 7T fMRI data while participants viewed a Ternus-Pikler display, where we pitted retinotopic vs. non-retinotopic motion processing. Three checkerboards moved horizontally back and forth. The checkerboards either did or did not alternate in contrast polarity with each frame. Because of the non-retinotopic reference frame, the percept was inverted such that the non-alternating checkerboards were perceived to alternate and vice versa. In V1, V2, and V3, the BOLD signal reflected retinotopic processing. In hMT+, activity correlated with the non-retinotopic percept. We propose that the motion of the reference frame (bike, checkerboards) is first detected in hMT+ and immediately subtracted from the motion of its parts, in line with our EEG results reflecting non-retinotopic processing from ~120 ms on.