However, in a video, adjacent frames share a large fraction of similar spatial information [4], especially the high-level semantics that alters gradually over temporal sequences in videos [5].