Training loss can not converge with custome cropped dataset.

[hot-dog]

Thanks for your great work! I tried with my own data and the training loss could not converge, all the training parameters are default. My data has been cropped and the principal point does not locate on the center of the image, i have modified the projection matrix accordingly and these modifications works fine with 3DGS, but do not work with 2DGS. I have no idea how to fix this, could you please give me some advises? Thank you:)

[hbb1]

Hi, we currently don't support non-ideal pinhole camera yet. Since we use a different projection method, modifying the projection matrix just doesn't work for 2DGS. To support this, we need to modify the rasterizer to support non-center intrinsic here https://github.com/hbb1/diff-surfel-rasterization/blob/main/cuda_rasterizer/forward.cu#L208. You can refer to this issue for making your data center.

[hot-dog]

Hi, we currently don't support non-ideal pinhole camera yet. Since we use a different projection method, modifying the projection matrix just doesn't work for 2DGS. To support this, we need to modify the rasterizer to support non-center intrinsic here https://github.com/hbb1/diff-surfel-rasterization/blob/main/cuda_rasterizer/forward.cu#L208. You can refer to this issue for making your data center.

Thanks for you reply, i get it, i will modify the rasterizer code and try again.

[hbb1]

Hi, can you check this PR #27 to see if the current implementation works for you?

BTW, make sure you recompile the latest rasterizer.

[hot-dog]

Sorry for the late reply, i modified the reaterizer code to adapt to principal point offset and the training can converge.