Learnable LOD bias

[ZhouCX117]

Thanks for your excellent work!
Could you please kindly refer to where I can find the learnable LOD bias in the code? Is it the "extra_level" parameter? Why this can supplement high-frequency regions?
And where is the code for smooth rendering transitions between different LOD levels without introducing visible artifacts?

[ZhouCX117]

@AnyiRao @tongji-rkr Sorry to bother you again. I wonder why the rotation parameter is set as unlearnable.

Octree-AnyGS/scene/gs_model_scaffoldgs/lod_model.py

Line 295 in eab6921

self._rotation = nn.Parameter(rots.requires_grad_(False))

[tongji-rkr]

First, we learn extra_level in the densification stage. The explanation for this problem is roughly as follows. We find that removing this parameter will cause LOD0 to fail to guarantee high-frequency regions (for example, there will be discontinuity at the texture and edge of the building). Therefore, we add extra_level to anchors with larger gradients so that they can be selected more easily.

[tongji-rkr]

Second, to smooth the transition, you need to set dist2level to progressive, as shown here

[tongji-rkr]

Third, this is a property of anchor, just to satisfy the prefilter process. The rotation of Gaussian is obtained by MLP,