Look at What I'm Doing: Self-Supervised Spatial Grounding of Narrations in Instructional Videos.