Here’s an interesting challenge. Can you train a CNN where, if you use the CAM heatmap technique from lesson 7, it highlights the dog in the image here?:
I suspect that our dogs v cats model or even basic imagenet model may be successful. I’d be interested if anyone makes it work!
But to be honest. I did cheat a bit. I searched for some google images with the keywords ‘Dalmatian’ and ‘Dalmatian Snow’ and added them to the training set (100 images from each search ) Of course, I visually inspected to ensure this particular image was not in the download list.
I prolly should’ve also tried with the existing cats/dogs image-set, but I didn’t. I also think solving this problem in the scope of the dog-breeding challenge might shed more insight on the model performance.