By default, SKImage enforces connectivity between superpixels. This leads to a much more coherent result for images that are noisy (e.g. handles leaves better). The major drawback of my algorithm is that I did not implement this. However, when I toggle the enforce_connectivity flag in the SKImage implementation, the results are comparable to that of my algorithm.
python main.py demo1 -i shrek.jpeg -a resnet152 -t layer4| Predicted Class | #1 mask | #3 neck brace |
|---|---|---|
| Grad-CAM |  |
 |
| Vanilla backpropagation |  |
 |
| Guided Grad-CAM |  |
 |
python main.py demo2 -i bull_mastiff.jpg
python main.py demo2 -i perturbed.jpg
| Layer | relu |
layer1 |
layer2 |
layer3 |
layer4 |
|---|---|---|---|---|---|
| Grad-CAM |  |
 |
 |
 |
 |
| Grad-CAM (After Perturbation) |  |
 |
 |
 |
 |