fix camera intrinsics

[shewangmu]

Hello, I find that in some situations, the rendering result of GS is very blurry, especially around the edge. This behaves very differently from other cases.
Then I locate the problem with camera intrinsics. You didn't consider about the cx and cy.
After this bug fixing, the result improves a lot.
Additionally, just a reminder, I find the L2 loss behaves more stable than L1. L1 seems to cause more artifacts (in object rendering).

[grgkopanas]

Indeed that is true, but we cannot accept this pull request. To be complete you need to make sure that you consider the cx and cy when computing the forward and backward pass in cuda when we compute the 2D covariance matrix of the Gaussians based on the jacobian of the transformation.

[shewangmu]

Indeed that is true, but we cannot accept this pull request. To be complete you need to make sure that you consider the cx and cy when computing the forward and backward pass in cuda when we compute the 2D covariance matrix of the Gaussians based on the jacobian of the transformation.

Thank you for your suggestion, I will check it on Monday. However, I test this change under several situations and it works right. I am not sure whether it’s necessary or not.

[kwea123]

just change the projection matrix to this #399 and it works with the current cuda implementation.

[shewangmu]

just change the projection matrix to this #399 and it works with the current cuda implementation.

Yes, the projection matrix in the issue you mentioned is just the same with this commit. And I have checked the result under many cases.

[kwea123]

it works correctly