Latest 3DGS Native Implementation

[widifadi]

Hi, I'm wondering if this current v2 repository is using the latest 3DGS improvement especially on adding sparse_adam acceleration and depth regularization? Also, I'm wondering if it is possible to use the training result from 3DGS native onto the masking part?

Thanks!

[widifadi]

I'm also wondering the difference between this repository and the one in https://github.com/Jumpat/SegmentAnythingin3D/tree/sa3d-gs

[Jumpat]

Hi, I'm wondering if this current v2 repository is using the latest 3DGS improvement especially on adding sparse_adam acceleration and depth regularization? Also, I'm wondering if it is possible to use the training result from 3DGS native onto the masking part?

Thanks!

Hello, sparse_adam acceleration and depth regularization do not affect the original 3D Gaussian representation, so SAGA can seamlessly integrate these modifications. However, our code does not use the latest version of 3D-GS. Nevertheless, the 3DGS training results can still be used for the mask part.

I'm also wondering the difference between this repository and the one in https://github.com/Jumpat/SegmentAnythingin3D/tree/sa3d-gs

The SA3D-GS is another tech for 3D-GS segmentation. Different from SAGA, it adopts an iterative pipeline for 3D segmentation. You can find the concrete method in our paper: https://link.springer.com/article/10.1007/s11263-025-02421-7

[widifadi]

Hi, I'm wondering if this current v2 repository is using the latest 3DGS improvement especially on adding sparse_adam acceleration and depth regularization? Also, I'm wondering if it is possible to use the training result from 3DGS native onto the masking part? Thanks!

Hello, sparse_adam acceleration and depth regularization do not affect the original 3D Gaussian representation, so SAGA can seamlessly integrate these modifications. However, our code does not use the latest version of 3D-GS. Nevertheless, the 3DGS training results can still be used for the mask part.

I'm also wondering the difference between this repository and the one in https://github.com/Jumpat/SegmentAnythingin3D/tree/sa3d-gs

The SA3D-GS is another tech for 3D-GS segmentation. Different from SAGA, it adopts an iterative pipeline for 3D segmentation. You can find the concrete method in our paper: https://link.springer.com/article/10.1007/s11263-025-02421-7

Thanks for the reply. I'll try using with the preprocessed data using the native 3DGS result on my next experiment then. Will check out on the other paper as well, thanks!