Discovering discriminative action parts from mid-level video representations | IEEE Conference Publication | IEEE Xplore