Are the pretrained model weights updated?

[littlepure2333]

Hi, thanks for your amazing work. However, I have a question. I ran the same inference code on the same image once in May and once yesterday, but I got quite different results (for example, depth). I'm wondering if you have updated the pretrained model weights (DUSt3R_ViTLarge_BaseDecoder_512_dpt)?

Image:

Before:

After:

[yocabon]

Hi, yes, all three checkpoints were updated in June. We made some changes to the training recipe (we swapped one dataset for another).

[littlepure2333]

Thanks for your prompt reply. May I know which dataset is swapped?

[puyiwen]

@littlepure2333 Hi, do you know the depthmap is metric depth or relative depth? Thank you!

[littlepure2333]

@puyiwen In Dust3r, it's scale-invariant depth. If you want metric depth, you can check out Mast3r.

[puyiwen]

@littlepure2333 Thank you for your reply! Do you run the Linear head model? I found that the Linear head seems to get metric depth, while the DPT head gets relative depth.

[littlepure2333]

@puyiwen Sorry I didn't run the liner head model. But since the training data is normalized in dust3r, all output depth of different dust3r variant should be scale-invariant.