Revisiting knowledge transfer for training object class detectors

  Abstract