Q 1.1.5
Rendered GIF from the 3D Gaussian Splatting rasterizer
Rendered GIF from the 3D Gaussian Splatting rasterizer
Learning rates used for each parameter
Means learning rate: 0.001
Covariance learning rate: 0.005
Opacity learning rate: 0.01
Color learning rate: 0.0025
Iterations: 1000
PSNR: 29.326
SSIM: 0.932
Comparison 1: The left image (without SH) shows a more uniform, flat appearance with limited color variation, while the right image (with SH) exhibits richer color gradients and more realistic lighting effects. The spherical harmonics allow the model to capture view-dependent color variations, resulting in more natural shading and highlights that better represent the 3D geometry of the object.
Comparison 2: This comparison highlights the improved directional lighting representation with spherical harmonics. The image with SH (right) demonstrates smoother color transitions and more accurate representation of how light interacts with different surface orientations. The non-SH version (left) appears more simplistic with less sophisticated color modeling, particularly noticeable in regions where the surface normal changes significantly relative to the viewing direction.
Without guidance
Iterations: 1000
With guidance
Iterations: 1000
Without guidance
With guidance
Without guidance
With guidance
Without guidance
With guidance
Color render
Depth
Color render
Depth
Color render
Depth
Color render
Depth
Color render
Depth
Color render
Depth