Next-active-object prediction from egocentric videos