Using body-anchored priors for identifying actions in single images