3D Gaussian parameters
All stored as Variables for gradient computation
Constructors
from_points(points : Tensor, colors : Tensor | Nil = nil, device : Tensor::Device = Tensor::Device::GPU) : Gaussian3D
Initialize from point cloud
new(count :
Int32, device :
Tensor::Device =
Tensor::Device::GPU)
SourceInstance methods
clone_gaussians(indices :
Array(
Int32)) :
Gaussian3D Clone Gaussians (for densification)
concat!(other : Gaussian3D) : Nil
Concatenate another Gaussian set
log_scale
Log-scale for covariance [N, 3]
Actual scale = exp(log_scale)
log_scale=(log_scale : Autograd::Variable)
Log-scale for covariance [N, 3]
Actual scale = exp(log_scale)
logit_opacity
Opacity in logit space [N, 1]
Actual opacity = sigmoid(logit_opacity)
logit_opacity=(logit_opacity : Autograd::Variable)
Opacity in logit space [N, 1]
Actual opacity = sigmoid(logit_opacity)
normalize_rotations!
Normalize quaternions to unit length
opacity
Get actual opacity values (sigmoid of logit_opacity)
parameters
Get all trainable parameters
position
Position in world space [N, 3]
position=(position : Autograd::Variable)
Position in world space [N, 3]
remove!(indices_to_remove :
Set(
Int32)) :
Nil Remove Gaussians by indices (for pruning)
reset_opacity!(value :
Float32 = -2.0_f32) :
Nil Reset opacity for all Gaussians (used periodically during training)
rotation
Rotation as quaternion [N, 4] (w, x, y, z)
Normalized to unit quaternion
rotation=(rotation : Autograd::Variable)
Rotation as quaternion [N, 4] (w, x, y, z)
Normalized to unit quaternion
scale
Get actual scale values (exp of log_scale)
sh_coeffs
Spherical harmonics coefficients [N, SH_COEFFS, 3]
For view-dependent color
sh_coeffs=(sh_coeffs : Autograd::Variable)
Spherical harmonics coefficients [N, SH_COEFFS, 3]
For view-dependent color