Computational Analysis Of Visual Motion PDF Download

We live in a phenomenally complex visual world, yet the human visual system extracts behaviorally meaningful information with apparent ease and efficiency. Our brains experience visual motion daily, and perhaps the most intricate and fascinating movement patterns are those of living creatures. Of particular importance are the actions of other humans, which contain rich information with social and biological relevance. The presented research investigates biological motion perception using relatively novel psychophysical techniques. In each of three experiments we use a unique variant of the "Bubbles" reverse correlation method, which works generally by revealing only portions of a stimulus randomly across many trials and then reverse correlating observer performance to illuminate the most informative regions of the stimulus. Experiment 1 uses "Temporal Bubbles", a new adaptation of "Bubbles" to the time domain, to determine if particular moments or postures during a point-light action sequence are more informative than others. Results show that there are indeed particularly diagnostic intervals in action sequences, but moments in this interval are not necessarily more informative if presented in isolation as static postures. We conclude that specific mid-level motion features are most critical for perceiving biological motion. In Experiment 2 we further elucidate these critical features by using "Spatio-temporal Bubbles" and quantitatively comparing human performance to a biologically inspired computational model of perception. Observers apparently use the same mid-level motion and form features for both point-light and stick figure sequences. Additionally, observer performance correlates with the "form pathway" of the model when stimulus duration is short (