search

zhuduowang / ChangeViT

The officical code of 'ChangeViT: Unleashing Plain Vision Transformers for Change Detection'.
Other
33 stars 4 forks source link

Input normalization #5

Closed kurbobo closed 3 weeks ago

kurbobo commented 1 month ago

Hi, thanks for your great work! It is very easy to work with! As far as I can see from the paper, you use pretrained dino-v2, but there is 0.5 input normalization in your code, that can corrupt the pretrained vit result Or maybe you retrained dino by yourself? If you use pretrained dino-v2 by Facebook, then did you tried dino-v2 with registers? Thank you in advance

zhuduowang commented 3 weeks ago

Thank you for your interest in our work. Adapting DINOv2 to remote sensing tasks and using 0.5 as the normalization value does not affect the pre-trained parameters, as we fine-tune the entire DINOv2 model. Additionally, we do not use DINOv2 with registers; this remains an area for future exploration.