Modeling Sense Disambiguation of Human Pose: Recognizing Action at a Distance by Key Poses