search

mhamilton723 / STEGO

Unsupervised Semantic Segmentation by Distilling Feature Correspondences
MIT License
711 stars 142 forks source link

Running STEGO with greyscale images. #65

Closed den-107 closed 1 year ago

den-107 commented 1 year ago

Hello, anyone had luck training stego with greyscale images? In my case it seems I cant get the needed pretrained weights for the DinoFeaturizer class. Anyone has a url to load these weights?

MohamadHawshar commented 1 year ago

Triple the channels of your images or change the line 101 in Data.py to open the images as RGB: img = Image.open(join(self.img_dir, image_fn)).convert('RGB')

mhamilton723 commented 1 year ago

Tha you for answering this @MohamadHawshar! Yes DINO pretrained only works with RGB images so you need to bump up the channels or make your on special DINO