Why abandon ZoeDepth in metric depth fine-tuning of V2?

[liboyu02]

As in title, the code of metric depth estimation in V2 doesn't seem to include ZoeDepth. But in V1, it is obvious to use ZoeDepth. Why this change?

[LiheYoung]

Hi @liboyu02, we continue to use ZoeDepth for generating our metric depth results in V2, as shown in Table 4. You can reproduce these results using the codebase from V1 by loading the pre-trained weights of V2. The metric depth estimation code in V2 is intended as a simpler alternative to the complex ZoeDepth codebase for training metric depth models. We implemented a basic DPT-based model using a straightforward regression-based loss function for this purpose.

[I3aer]

Hi @liboyu02, we continue to use ZoeDepth for generating our metric depth results in V2, as shown in Table 4. You can reproduce these results using the codebase from V1 by loading the pre-trained weights of V2. The metric depth estimation code in V2 is intended as a simpler alternative to the complex ZoeDepth codebase for training metric depth models. We implemented a basic DPT-based model using a straightforward regression-based loss function for this purpose.

Can you elaborate it further? Do you add another head on DPT or use the available head for relative depth for fine-tuning?