Request for Guidance on Training Data Generation Pipeline

[ycWang9725]

Dear UniDepth Team,

Thank you for your incredible work on UniDepth (both v1 and v2)! The methodological rigor and impressive results have been truly inspiring. Your contributions have motivated me to dive deeper into this field and build upon your achievements.

However, I've run into some challenges while trying to replicate the full training setup described in the v2 paper - specifically regarding the generation of the complete 16M training dataset. While I've tried to process the open-source raw resources, aligning depth maps, 3D points, and camera parameters with your pipeline requires substantial engineering effort that currently exceeds my resources.

Would your team consider sharing either:

1. The pre-processed training/test data in HDF5 format, or
2. The data transformation code used to convert raw datasets into your training format?

I completely understand if this isn't feasible due to technical constraints or proprietary considerations - I recognize the dataset represents years of careful curation. If direct sharing isn't possible, might there be additional implementation details about dataset generation that you could share beyond what's in the paper?

Either way, thank you for open-sourcing your work and pushing the field forward. Your transparency has already been tremendously valuable to the community. Please let me know if I can clarify this request further.

[DrawingProcess]

I'm facing the same issue — would it be possible to share the preprocessing code or more details on the dataset generation? Thanks!

[dxw2000]

I have the same need. Has anyone solved it?