How to generate rays_<datasetname>.dat to train

[LuckyOne09]

Hello, thank you for your great and complete work first. But I come across a problem when try to reconstruct a scene by ngp.
I am reconstructing a scene captured by cameras with different intrinsic parameters. But I notice ngp assumes using same camera in transform.json.
Following this hint , I want to pre-compute all the rays, but I dont know how to generate that, could you give me some guidelines?

[Tom94]

Hi there, you can actually provide per-frame intrinsics by now -- so you don't need to precompute rays. Simply bundle the same parameters that are global by default ("cx", "cy", etc.) within the frame's scope (alongside "file_path" and so on). All frames with such custom intrinsics will use those instead of the global ones. Cheers!

[Tom94]

Yes

[fishfishson]

Hi @Tom94 I find in run.py file, ngp uses global fov to render images. Can I use per frame intrinsic parameters like parsing the above json file to render images? If so, could you pls tell me how to implement it?

[Tom94]

You can use the Python bindings (by modifying run.py) to change testbed's FoV setting depending on which image you're rendering out. See python_api.cu for all the available bindings (you should be able to find the ones controlling fov there).

[fishfishson]

like testbed.set_fov_xy = [angle_x, angle_y]?

[Mabel0403]

like testbed.set_fov_xy = [angle_x, angle_y]?

Hi @fishfishson, I have similar problems as you. I used images with different intrinsics, getting a NeRF model and base.ingp. Then, I tried to render an image by giving a .json. However, the image which run.py outputed isn't from the perspective I need. Can I discuss this problem with you in detail? Thanks a lot!