Figure 6
EigenCAM59 activation maps on several images for our four different models. The images in the top and bottom row are from positive and negative classes respectively. The DINO-ViT backbone addresses similar areas to those proposed by human annotated DogFACS variables.
