About Table 6 in the Original Paper

[zhangzw12319]

Thanks for your excellent and outstanding efforts in this work! Really love the work.

As for the transferring performance experiments in Table 6, I have some questions about the detailed training settings:

• Is either ViTS, ViTB or ViTL used for this experiment?
• Is this model trained from scratch (using DinoV2 original weights and finetuning on speific single dataset) in Table 6, or trained by self-supervised manner on the whole labeled 1.5M and 65M unlabeled dataset? In other words, is there self-supervision manner involved in this experiment settings?

THX a lot for your kind reply and description.

[LiheYoung]

Thank you for your questions.

1. ViT-L is used.
2. The model is fine-tuned from DINOv2 weights, without further performing SSL on our 1.5M + 62M depth data.

[zhangzw12319]

Thx for your kind information !