Learning gestures from partial ground truth