Question About Supporting Images with Different Intrinsics

[Rashfu]

The context images used in the model seem to have the same intrinsic during the training process. Do your pre-trained weights support inputting two images with different intrinsics for training or inference? If not, do I need to retrain the model?

Thank you in advance for your time!

[botaoye]

Hello, thank you for your insightful question. We train directly with re10k and dl3dv, so all image pairs have the same intrinsics. I have not tested the case where the input image pairs have different intrinsics, but I think retraining on data with different intrinsics gives better performance. If you have any results and would like to share them, that would be great.

[Rashfu]

Thanks for your explanation. Based on my tests, the current pre-trained model doesn't support different intrinsics. Therefore, it needs to be retrained on data with different intrinsics.