Parameters for seeding Gaussians from a ReUseX point cloud. More...

#include <GaussianCloud.hpp>

Public Attributes
int	knn = 3
	Neighbours used for the nearest-neighbour spacing that sets the initial (isotropic) scale of each Gaussian.
float	initial_opacity = 0.1f
	Initial alpha before the logit transform.
float	min_scale = 1e-4f
	Clamp on the initial scale [m].
float	max_scale = 0.5f
int	sh_degree = 0
	Active SH degree.
std::size_t	max_points = 0
	Uniform stride cap on the number of seeds (0 = use every point).

Detailed Description

Parameters for seeding Gaussians from a ReUseX point cloud.

Definition at line 65 of file GaussianCloud.hpp.

Member Data Documentation

◆ initial_opacity

float reusex::gsplat::GaussianInitOptions::initial_opacity = 0.1f

Initial alpha before the logit transform.

Definition at line 72 of file GaussianCloud.hpp.

◆ knn

int reusex::gsplat::GaussianInitOptions::knn = 3

Neighbours used for the nearest-neighbour spacing that sets the initial (isotropic) scale of each Gaussian.

The reference implementation uses the mean squared distance to the 3 nearest points.

Definition at line 69 of file GaussianCloud.hpp.

◆ max_points

std::size_t reusex::gsplat::GaussianInitOptions::max_points = 0

Uniform stride cap on the number of seeds (0 = use every point).

There is deliberately no RNG seed here: seeding picks every stride-th point, which is reproducible without consuming randomness at all (STANDARDS §6). TrainOptions::seed covers the parts that are random.

Definition at line 89 of file GaussianCloud.hpp.

◆ max_scale

float reusex::gsplat::GaussianInitOptions::max_scale = 0.5f

Definition at line 78 of file GaussianCloud.hpp.

◆ min_scale

float reusex::gsplat::GaussianInitOptions::min_scale = 1e-4f

Clamp on the initial scale [m].

A LiDAR cloud can contain near-duplicate points whose neighbour distance is ~0; without a floor those Gaussians start infinitely small and never receive gradient.

Definition at line 77 of file GaussianCloud.hpp.

◆ sh_degree

int reusex::gsplat::GaussianInitOptions::sh_degree = 0

Active SH degree.

0 keeps only the DC (view-independent) colour, which is what a short training run can actually fit.

Definition at line 82 of file GaussianCloud.hpp.

The documentation for this struct was generated from the following file:

libs/reusex/include/gsplat/GaussianCloud.hpp

Public Attributes