Although the bounding box supervision is relatively more costly than the simple image-level class labels, the proposed method exploits the benefit of the given annotation effectively and achieves good performance. In particular, the MIL formulation is interesting. I recommend accepting this paper.