Excellent work, confusion about the code

[BINOFL]

I have read your scaffold-gs and octree-gs code, and for this project, this is six different types of gaussian models, which are respectively [2dgs/3dgs/scaffold-gs] + [basic/lod]. Can you make your readme.md more specific for these six models?

Like what's the difference of them, i quickly debug and read this code with [3dgs + basic], but i found it use voxel to control the scene which i think might only appear in [scaffold-gs + basic/lod]. Thanks

[tongji-rkr]

First, we support both the general and LOD versions. For the generic version, all Gaussians (explicit Gaussians or neural Gaussians) are bound to anchors and use the similar densification strategy in Scaffold-GS.

[tongji-rkr]

I also explored the strategy of implementing LOD based on the vanilla 3DGS, but encountered some problems in the strategy of densification, and the final effect was slightly lower than that of 3DGS. In addition, through the unified densification strategy, the generality of the code is also improved

[tongji-rkr]

In addition, we plan to update the code at the end of November and add details such as depth supervision and transforming 3DGS

[BINOFL]

I understand, so as you explained, what is the different when i use 3dgs+basic and scaffold-gs+basic ? sounds like bounding 3DGS to voxels and use the same scaffoldgs optimisation methods makes it hardly no difference with the scaffold-gs method.

[jianglh-WHU]

whether or not use MLP to decode the attributes of Gaussian primitives.