Does pinhole‑camera images are supported

[wubinwab]

Hello, could you please confirm whether the released code and weights support Gaussian Splatting reconstruction from pinhole‑camera images?
I have conducted experiments using the front‑view data from Waymo. I implemented the ray‑generation module on my own, yet the reconstructed point cloud quality is poor.
Therefore, I would like to verify if your provided weights are compatible with pinhole‑camera inputs.
Thank you!

[daehyoungko]

Thanks for the report.

In the current released inference path, raw OpenCV pinhole cameras are not directly supported.
The camera conversion and ray-generation code are intentionally F-theta-only.
the remaining “simple pinhole” conversion is used internally for FOV conditioning and is not a supported input camera model.

[wubinwab]

Thanks for your reply.
And I implemented the ray‑generation module on my own, yet the reconstructed point cloud quality is poor on Waymo front view image. Does the model checkpoint released is trained based on 120 degree F-theta images?

[Enamel2]

I tried waymo data and recalibrated the images to f-theta model, the error was sub-pixel, and also found the result is very blurred and hazy.
I also done another experiment by cropping the image from front-left and front-right cam on Nvidia dataset by about 10 deg of FOV, then found that the depth inference result degenerate a lot. So I more believe that current released model is kind of overfitted to Nvidia's own rig. Migration to other datasets might need fine tuning.

[heiwang1997]

Hi you could refer to #9 for more discussion on this matter.