This demo shows the method proposed in "Zhou, Bolei, et al. "Learning Deep Features for Discriminative Localization." arXiv preprint arXiv:1512.04150 (2015)".
The proposed method can automatically localize the discriminative regions in an image using global average pooling (GAP) in CNNs.
You can download the pretrained Inception-V3 network from here. Other networks with similar structure(use global average pooling after the last conv feature map) should also work.