Awesome

There and Back Again: Revisiting Backpropagation Saliency Methods

Standalone code for the paper: There and Back Again: Revisiting Backpropagation Saliency Methods, CVPR 2020 by Sylvestre-Alvise Rebuffi*, Ruth Fong*, Xu Ji* and Andrea Vedaldi.

For better analysis tools and to perform the benchmarks like the Pointing Game, we recommend to use TorchRay where all the saliency methods of the paper are implemented.

<img src="splash.gif"> Combinations of phase 1 (rows) and phase 2 (columns) for VGG16 at different layers ("combi" is when combining layers)

Code

To get the saliency maps used in the above figure, simply run the following PyTorch code:

python vgg16_grid_saliency.py

The other python file produces the same grid but for ResNet50.

Cite this work

If you use this code for your project please consider citing us:

@inproceedings{rebuffi2020saliency,
  title={There and Back Again: Revisiting Backpropagation Saliency Methods},
  author={Rebuffi, Sylvestre-Alvise and Fong, Ruth and Ji, Xu and Vedaldi, Andrea},
  booktitle={Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2020}
}

Acknowledgments

This work is supported by Mathworks/DTA DFR02620, Open Philanthropy, EPSRC AIMS CDT and ERC 638009-IDIU.

The standalone code in this repository is inspired by https://github.com/jacobgil/pytorch-grad-cam