Code for the paper:

Oral at BMVC 2020,
Authors:
Ruigang Fu, Qingyong Hu, Xiaohu Dong, Yulan Guo, Yinghui Gao and Biao Li,

XGrad-CAM is a CNN visualization method, try to explain why classification CNNs predict what they predict. It is class-discriminative, efficient and able to highlight the regions belonging to the objects of interest.

The main difference between XGrad-CAM and Grad-CAM locates at line 116 - line120:

X_weights = np.sum(grads_val[0, :] * target, axis=(1, 2))

X_weights = X_weights / (np.sum(target, axis=(1, 2)) + 1e-6)

weights = np.mean(grads_val, axis=(2, 3))[0, :]

Usage: python XGrad-cam.py --image-path <path_to_image> --target-index <class_of_interest> for CPU computation, add --use-cuda for GPU acceleration.

Example: python XGrad-CAM.py --image-path ./examples/ILSVRC2012_val_00000077.JPEG --target-index 159 --use-cuda

Output: class of interest: n02087394 Rhodesian ridgeback

Results:

left is Grad-CAM, right is XGrad-CAM

This is a simple script of experimental proof for our statement that given an arbitrary layer in ReLU-CNNs, there exists a specific equation between the class score and the feature maps of the layer (Eq.(5) in our paper).

Usage: python Proof_verify.py --image-path <path_to_image> --target-index <class_of_interest>

For any class of interest, the result will show that class_score-gradients*feature-bias_term=0

These codes are based on https://github.com/jacobgil/pytorch-grad-cam. Thanks to the author Jacob Gildenblat for the beautiful original code.

If these codes are useful to you, please cite our work:

@misc{fu2020axiombased,
    title={Axiom-based Grad-CAM: Towards Accurate Visualization and Explanation of CNNs},
    author={Ruigang Fu and Qingyong Hu and Xiaohu Dong and Yulan Guo and Yinghui Gao and Biao Li},
    year={2020},
    eprint={2008.02312},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}