FusionSeg: Learning to combine motion and appearance for fully automatic segmention of generic objects in videos.