Tube ConvNets: Better exploiting motion for action recognition | IEEE Conference Publication | IEEE Xplore