Is it possible to train a Gaussian model using data from a pinhole camera and then render it in fisheye camera mode?

[bolunzhangQwQ]

If possible, how should it be implemented specifically? Simply changing the is_fisheye parameter in the render doesn't seem to achieve it. Thank you for your valuable time.

[zmliao]

It is possible but some bugs are reported when training on pinhole camera model. We will release a debugged version later.

If you want to train a Gaussian model from a certain camera model, use the command below:
python train.py
-s
-m
--images
--colmaps
--bs 3
--ds 1
--camera_model <FISHEYE/PINHOLE>
--train_random_background

[zmliao]

Config the arguments to change your camera model when training or rendering.

[bolunzhangQwQ]

I'll try later. Looking forward to your code, thank you.

[zmliao]

I'll try later. Looking forward to your code, thank you.

I have found a bug, that the focals of the pinhole camera are different from the fisheye camera. Try to load the intrinsics using /dlsr/nerfstudio/transforms_undistorted.json when training from pinhole cameras. I will release the code after debugging the dataloader.