Closed Alexanderisgod closed 10 months ago
Given that MaskCut requires the application of normalized cuts to the patch-wise affinity matrix to produce segmentation masks, utilizing a ViT-based backbone becomes imperative. Moreover, the emerging segmentation properties are observed in self-supervised vision transformers, as detailed in the DINO paper. Consequently, MaskCut needs to harness the capabilities of the self-supervised ViT model for effective unsupervised segmentation.
when i replace the Vit backbone to resnet, maskcut result uncorrect mask, so the backbone can only use Transformer ways?