Optical snow

When an observer views a cluttered 3D scene,  visible objects in the scene undergo motion parallax.  Objects that are closer to the observer move faster than than those that are far away.  Because cluttered scenes produce a large number of occlusions and depth discontinuities, the resulting image velocity field also contains a large number of discontinuities.   These discontinuities pose a challenge for the visual system, which needs to simultaneously estimate image velocities and velocity field discontuities.  In particular, classical methods for measuring image velocities assume the motion the scene consists of a smooth surface or a small number (usually two) of depth layers, but this assumption fails miserably in 3D cluttered scenes.  We referred to such complex motion fields as optical snow, since the massively discontinuous motion fields are similar to what a (static) observer sees when viewing falling snow.  We have addressed this problem by developing a set of techniques which are based on a 3D frequency domain analysis (power spectrum of XYT).

Main Publications