Awesome

Group-CAM

By Zhang, Qinglong and Rao, Lu and Yang, Yubin

[State Key Laboratory for Novel Software Technology at Nanjing University]

This repo is the official implementation of "Group-CAM: Group Score-Weighted Visual Explanations for Deep Convolutional Networks".

Updates

2021/08/18 *

Adding Grad-CAM, Guided_BP, IG, RISE, Score-CAM and Smooth Grad
Cluster methods for grouping are supported in Group-CAM
Adding demo for insertion and deletion

Approach

<div align="center"> <img src="https://github.com/wofmanaf/Group-CAM/blob/master/figure/fig_1.png"> </div> <p align="center"> Figure 1: Pipeline of Group-CAM. </p>

Target layer

ResNet: 'layer4.2', Vgg19: 'features.35'

Demo

To visualize a heatmap, run:

python demo.py --arch vgg19 --target_layer features.35 --input Images/ILSVRC2012_val_00000073.JPEG --output base.png

with insertion and deletion curves:

python demo.py --arch vgg19 --target_layer features.35 --input Images/ILSVRC2012_val_00000073.JPEG --output base.png --ins_del

Citing Group-CAM

@article{zhql2021gc,
  title={Group-CAM: Group Score-Weighted Visual Explanations for Deep Convolutional Networks},
  author={Zhang, Qinglong and Rao, Lu and Yang, Yubin},
  journal={arXiv preprint arXiv:2103.13859},
  year={2021}
}