Downstream task reproduce

[jo1jun]

Hello, thank you for the excellent research and contributions.

I have some questions regarding downstream task training and evaluation.

Using Colab, I performed depth estimation on KITTI and NYU datasets with the DINOv2 pretrained ViT-S/14 and ViT-B/14 original models, as well as with the FiT3D fine-tuned models. I followed the repositories provided at DINOv2 and Monocular-Depth-Estimation-Toolbox.

I trained with the same configuration as in the paper for comparison.

While the RMSE values from the original model closely match those in the paper:

• ViT-S/14 (KITTI: 3.0721, NYU: 0.4416)
• ViT-B/14 (KITTI: 2.9617, NYU: 0.3999)

the FiT3D fine-tuned models show performance below what is reported:

• ViT-S/14 (KITTI: 3.6037, NYU: 0.5057)
• ViT-B/14 (KITTI: 3.1444, NYU: 0.4907)

Is there something I might have missed during reproduction? The weights appear to load correctly, and I used the Colab pre-load model section directly.

---

Additionally, could you share any details on how you configured the ADE20k and Pascal VOC downstream tasks or if there are plans to release downstream task training/evaluation code?

I am considering using the mmsegmentation repository, similar to how I used the MDE toolbox.

Thank you for your time and help!

[ywyue]

Hi @jo1jun, thanks for your interest in our work.

In the Colab demo and all the visualizations of feature maps and k-means clustering in the paper, we only use the fine-tuned features. However, for experiments on linear probing evaluation, we combine the fine-tuned features with the original features. As the 2D models were only fine-tuned on a small-scale indoor dataset ScanNet++, their generalization may degrade, which is one limitation of our work. We found that simply concatenating the original 2D features with fine-tuned features can preserve their generalization while incorporating 3D awareness. Related discussion can be found in:

• Feature assembly in Sec. 3.3
• Limitations and discussion in Sec. 4.6
• Appendix C. Experiments on Feature Dimensions

We conducted the ADE20k and Pascal VOC segmentation tasks using the mmsegmentation library. For setup, please refer to DINOv2's config:

• ViT-B/14:
  • ADE20k: https://dl.fbaipublicfiles.com/dinov2/dinov2_vitb14/dinov2_vitb14_ade20k_linear_config.py
  • Pascal VOC: https://dl.fbaipublicfiles.com/dinov2/dinov2_vitb14/dinov2_vitb14_voc2012_linear_config.py
• ViT-S/14:
  • ADE20k: https://dl.fbaipublicfiles.com/dinov2/dinov2_vits14/dinov2_vits14_ade20k_linear_config.py
  • Pascal VOC: https://dl.fbaipublicfiles.com/dinov2/dinov2_vits14/dinov2_vits14_voc2012_linear_config.py

I am cleaning the linear probing evaluation code and will try to release it in this week.

[jo1jun]

I realized I missed the implementation details in the paper! Thank you for your kind and helpful response! I’m looking forward to the evaluation code. :)

[ywyue]

Hi @jo1jun, the linear probing evaluation code has been released. Sorry for the delay.