Hi Khaled, I'm doing my Msc thesis on combating spurious features in medical imaging, and your paper is a core paper serving as an essential stepping stone for my next steps. I have troubles replicating the plot with the improving Robust AUROC with increasingly higher values of spatial specificity. My models get the best Robust AUROC performance on the lower 10% quartile of spatial specificity, and as the bounding boxes gets tighter, they drop to random chance and a Robut AUROC on about 0.5. Any chance you can share your code for the synthetic experiment as well?
Much appreciated <3
Hi Khaled, I'm doing my Msc thesis on combating spurious features in medical imaging, and your paper is a core paper serving as an essential stepping stone for my next steps. I have troubles replicating the plot with the improving Robust AUROC with increasingly higher values of spatial specificity. My models get the best Robust AUROC performance on the lower 10% quartile of spatial specificity, and as the bounding boxes gets tighter, they drop to random chance and a Robut AUROC on about 0.5. Any chance you can share your code for the synthetic experiment as well?
Much appreciated <3