why not just use Gradient-base to gain a saliency may
zhaoxinjie opened this issue · 3 comments
Maybe I don't hnow clearly about this paper. but since we can get a saliency map just by gradient-base, why I get a CAM first, then use Guided-Backprop. Isn't just use gradient better? looking forward to your answer. thank you
The visualizations obtained by taking the gradients wrt the input pixels for any target class is not class discriminative. See figure 2 of the grad cam paper for an example. In our work we compute the gradients with respect to the convolution layer and do a weighted combination to get Grad-CAM (coarse heatmap). We can combine with this with other pixel level visualizations that are high resolution (but not class discriminative such as Guided Backprop) to get a high resolution class discriminative visualization-Guided Grad-CAM. Please let me know if it clears your doubt.
By CAM, I think you mean Grad-CAM. The Grad in Grad-CAM refers to the gradient they we compute wrt the last convolution layer
thank for your answer，i'm clear now.