Clarification on organizing HS-NeRF BaySpec data for training

[ming053l]

Hi, thank you for sharing the HS-NeRF dataset and your work.

I’m currently trying to use the BaySpec subset from the HS-NeRF dataset for training your model. However, the data structure is quite unclear and lacks documentation, which makes it difficult to interpret.

For example, in the BaySpec folder, there are many subfolder, and various other files, but it’s not obvious:

• Which files correspond to hyperspectral image cubes across different view
• Where the RGB images are located
• How to obtain or reconstruct camera pose information (transforms.json, intrinsics/extrinsics, etc.)
• How these files should be structured to be compatible with your codebase (e.g., expected directory structure like images/, hyperspectral/, transforms.json, etc.)

Could you kindly provide a short explanation or example of how to organize the BaySpec data for model training?

It would be very helpful to have at least one fully structured example (even minimal) to follow.

Thank you in advance!

[Factral]

Hi!, during development we had a lot of problems with that dataset because it’s quite disorganized. Processing it took us a while. I’d like to share the already processed files with you, but I currently don’t have access to the data.

You can open the files with the spectral library (each cube’s metadata is in the .hdr). The RGB images can be generated with the function in spec2rgb (you need to change the wavelength range and the gamma) and convert them to npy. They already provide a transforms.json; the only thing you need to do is adapt the new keys and update the file paths.

I’ll try to get access to the processed data again, and if I can, I’ll upload it to Drive.