Download docs/gaussian_math.md from SiddhuG01/PanoGS: direct link, hf CLI and curl.
- Browser
- Download file 3.48 kB
-
https://huggingface.co/spaces/SiddhuG01/PanoGS/resolve/main/docs/gaussian_math.md
- Command line
-
hf download hf://spaces/SiddhuG01/PanoGS/docs/gaussian_math.md
-
curl -L -o gaussian_math.md https://huggingface.co/spaces/SiddhuG01/PanoGS/resolve/main/docs/gaussian_math.md
A newer version of the Gradio SDK is available: 6.32.0
3D Gaussian Mathematics & Representation
This document details the mathematical formulation, parameterization spaces, 3D covariance computation, and spherical harmonics representation used in PanoGS.
1. 3D Gaussian Definition
A 3D Gaussian in world space is defined by its center mean $\mathbf{\mu} \in \mathbb{R}^3$ and a $3 \times 3$ symmetric positive semi-definite covariance matrix $\mathbf{\Sigma} \in \mathbb{R}^{3 \times 3}$:
2. Numerically Stable Parameterization
To ensure gradient updates remain in valid physical domains, Gaussians are represented using unconstrained internal parameters:
A. Scale $\mathbf{s} \in \mathbb{R}^3$ (Log-Space)
- Internal parameter: $\mathbf{s}_{\log} = (\ln s_x, \ln s_y, \ln s_z)$
- Physical scale: $\mathbf{s} = \exp(\mathbf{s}_{\log}) > 0$
B. Rotation $\mathbf{q} \in \mathbb{H}$ (Unit Quaternions)
- Internal quaternion: $\mathbf{q} = (q_w, q_x, q_y, q_z)$
- Normalized unit quaternion: $\tilde{\mathbf{q}} = \frac{\mathbf{q}}{|\mathbf{q}|_2}$
- Rotation matrix $\mathbf{R} \in \text{SO}(3)$: $$\mathbf{R} = \begin{bmatrix} 1 - 2(y^2 + z^2) & 2(xy - rz) & 2(xz + ry) \ 2(xy + rz) & 1 - 2(x^2 + z^2) & 2(yz - rx) \ 2(xz - ry) & 2(yz + rx) & 1 - 2(x^2 + y^2) \end{bmatrix}$$ where $(r, x, y, z) = (\tilde{q}_w, \tilde{q}_x, \tilde{q}_y, \tilde{q}_z)$.
C. Opacity $\alpha \in (0, 1)$ (Logit-Space)
- Internal parameter: $o_{\text{logit}} = \ln \frac{\alpha}{1 - \alpha}$
- Physical opacity: $\alpha = \sigma(o_{\text{logit}}) = \frac{1}{1 + \exp(-o_{\text{logit}})}$
D. Color Representation (Linear-Radiance Spherical Harmonics $SH_0$)
- Zeroth-order constant: $C_0 = \frac{1}{2\sqrt{\pi}} \approx 0.28209479177387814$
- RGB to $SH_0$: $\mathbf{f}_{\text{dc}} = \frac{\text{RGB} - 0.5}{C_0}$, where RGB is linear-light radiance.
- $SH_0$ to RGB: $\text{RGB} = \max(\mathbf{f}_{\text{dc}} \cdot C_0 + 0.5, 0.0)$. Values above 1.0 are retained by the float PLY representation for HDR.
- The legacy 32-byte
.splatformat stores 8-bit color and clips radiance to the display range. PanoGS.hdrsplatand Gaussian PLY preserve float radiance; the viewer converts HDR color to display sRGB and its exposure control sets the visible highlight range. .hdrsplatis PanoGS-specific (44 bytes per splat: XYZ, scale, float RGB radiance, float opacity, packed rotation); use.splatwhen interoperability with other viewers is required.
3. 3D Covariance Matrix Formulation
The 3D covariance matrix $\mathbf{\Sigma}$ is parameterized via scale $\mathbf{S} = \text{diag}(s_x, s_y, s_z)$ and rotation $\mathbf{R}$:
This matrix decomposition guarantees that $\mathbf{\Sigma}$ is symmetric and positive semi-definite ($\forall \mathbf{v} \neq 0, \mathbf{v}^T \mathbf{\Sigma} \mathbf{v} \ge 0$).
4. Adaptive Initialization from Point Clouds
Rather than assigning an arbitrary constant scale to all Gaussians, PanoGS initializes each Gaussian's scale $s_i$ dynamically based on the average Euclidean distance to its $k=3$ nearest spatial neighbors via cKDTree: