Training pipeline of linear probing for semantic segmentation and depth estimation

[ywyue]

Dear DINOv2 team, thank you for this amazing work! If I am correct, I only found the whole pipeline of linear probing for classification on ImageNet in https://github.com/facebookresearch/dinov2/blob/main/dinov2/eval/linear.py.

Is it possible for you to also release the training pipeline of linear probing for semantic segmentation (ADE20k, Cityscapes, Pascal VOC 2012) and depth estimation (NYU-Depth V2, KITTI, SUN-RGBD), including dataloader, data augmentation used, training schedule, etc? I believe this will serve as a standard testbed for follow-up works to compare with DINOv2 in a fair setup (i.e. excluding the effect of data augmentation, training schedule, etc). Thank you very much!

[zshn25]

Hi, here's an unofficial training and testing pipeline of segmentation using mmsegmentation, based on the configs provided by the authors. The example implements Cityscapes but can be adopted to other datasets

https://github.com/zshn25/DINOv2_mmseg