Abstract
The computational sub-goal of early visual motion processing is to estimate the spatial map of local image motion vectors (optical flow). Due to technical difficulty, little attempts have been made to psychophysically measure the optical flow experienced by human observers. Using a novel measurement method, the present study psychophysically estimated the perceived optical flow of naturalistic scenes. The target stimulus was a brief movie presented in a circular aperture. In the middle of the movie presentation, a small dot was flashed at the aperture center. The observer had to report the image motion vector at the location/timing of the flash by matching the speed and direction of the following matching stimulus (pink-noise field). The target and matching stimuli were alternatively presented several times until the observer was satisfied (method of adjustment). A preliminary experiment using a random-dot kinematogram as the target stimulus assured the reliability of our method (R2 =.88). We then applied this new method to more naturalistic stimuli, i.e., five short movie clips selected from the slow flow version of MPI Sintel Flow Dataset (an open-source 3D animated short film with grand-truth optical flow data). The perceived motion vector was measured at 36 locations (6 x 6 grid) for each clip. The directional bias was reduced by flipping each clip vertically, horizontally, or both. The perceived vector agreed to some extent with the local ground-truth vector at the flash point (R2=0.66), as well as with the ground-truth vector spatially averaged around the flash point by the best-fit Gaussian operator (R2=0.67). Nevertheless, at several locations (most of them were close to the object boundary), the perceived vector was significantly deviated from the local or averaged ground-truth vector. We have developed a promising method to reveal the characteristics of the human visual motion perception for naturalistic complex scenes.