Skip to content

DINOv2 Evaluation Results Different from the Paper #1

@LuneExplorer

Description

@LuneExplorer

Hello, thank you for your great work!
I have encountered some trouble when trying to reproduce your code. I can't reach the accuracy you reported in the paper. I used the the optimal parameters (layer=30 and C=1.0) you mentioned in the article to run the DINOv2 model on the depth task, but the accuracy I obtained is 87%, not 99%.
This is how I extract the DINOv2 features:

Image

Also, I visualized the region masks I generated using the NYUv2 toolbox and didn’t find any issues. The region IDs are also consistent with the region pairs file you provided.
So is there something I have done improperly?
Thank you.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions