Error with gradients when using camera with arbitrary principal points outside of Image Boundaries

[wagnva]

Hello,

I am currently trying to evaluate the use case of Gaussian Splatting for 3D-Reconstruction based on Satellite Imagery. Unfortunately the process used to approximate the satellite specific camera model into the required pinhole camera results in intrinsic camera parameters with arbitrary principal points, that can far exceed the size of the image.

One example camera intrinsics to show the large values:

After modifying the getProjectionMatrix to include the shift as described in issue #144 the initial projection of the SfM pointcloud can be seen correctly. Unfortunately the next few training iterations result in an unwanted shift of the gaussians outside of the camera field of view.

Timeseries of Training Iterations (Iteration 1 -> 5)

Has anyone made a similiar observation and/or has worked on modifying the code to handle arbitrary Principal Points correctly?