Can you release the code for DINOV2 alignment loss which is showed in 'Main Results'?

[bobojiang26]

[CIntellifusion]

We will try to release it if time permit. But you could directly replace VGGT model with DinoV2 to get a comparable result

[bobojiang26]

Thank you for reply! I still look forward to your release because the size of the feature extracted by DinoV2 is different from that of VGGT. The features of VGGT and DFoT are computed similarity on 24512512. Does DinoV2 feature alignment also do on this size? Would the size impact the performence?

[CIntellifusion]

In fact, we don't control the feature size [24,512,512] to be the same, the dinov2 has a much samller feature dimension. We just apply a same projector with different parameter to project dinov2 feature to the same size with dfot feature. For VGGT, we did try to interpolate the final latent to [24,256,256] which decreases the performance. However, the size for dinov2 feature is not sutdied.

[bobojiang26]

I see. Your reply really helps!