Open jdalch opened 2 months ago
In order to extract interesting concepts, we should use more concepts than classis, and divide the images into patches, like they do in craft. If we do so, the OOD score for the whole image should be calculated by aggregating the OOD scores per patch, e.g. by computing the mean of the patch-wise OOD scores. This experiment should be performed for the OpenOOD Imagenet-200 benchmark so that we can see the interesting concepts emerge.
Pour la suite du stage: