D-RISE with bad model

[QuentinAndre11]

Since D-RISE is archived, I cannot post an issue on this repo, but I actually think it could happen to have the same issue with RISE, I apologize if it does not suit. So, I have a bad model (with a very small recall, around 0.1): when the mask is applied, I get very few labels and thus the saliency map remains at 0 often. I guess I could decrease the probability of the masked superpixels but is it a good idea to decrease it too much ? I still have pretty bad results with 0.3. I also decreased the detect threshold of the model (I have no predicted labels for the masks over 0.2, and I start to have a correct percentage of null saliency maps at 0.08... but is it a good idea ?). Anyways, thanks a lot for your work!

[eclique]

Yes, getting meaningful saliency maps for models with bad accuracy/recall can be challenging. I don't see a problem with using low occlusion probability -- the intuition of the approach still works. For example, setting the probability to 1/(h*w) would be similar to the sliding window occlusion where one super-pixel is masked at a time.