Is the DINOv2 ViT-G encoder un-frozen when training teacher model?

DepthAnything / Depth-Anything-V2

[NeurIPS 2024] Depth Anything V2. A More Capable Foundation Model for Monocular Depth Estimation

https://depth-anything-v2.github.io

Apache License 2.0

3.86k stars 336 forks source link

Is the DINOv2 ViT-G encoder un-frozen when training teacher model? #37

Closed Nik-V9 closed 4 months ago

Nik-V9 commented 4 months ago

Hey folks, Thanks for the amazing work!

I was reading the paper regarding details for the training of the teacher model. It was not clear if the DINOv2 encoder is unfrozen when training the teacher model on the synthetic datasets?

Furthermore, similar to Depth Anything V1, is the DINOv2 encoder finetuned during the teacher-student model distillation?

Looking forward to your clarification!

LiheYoung commented 4 months ago

Hi, all DINOv2 models are fully fine-tuned in both V1 and V2.

Nik-V9 commented 4 months ago

Thanks for the clarification!