Assignment 4
Neural Volume Rendering and Surface Rendering
A. Neural Volume Rendering (80 points)
0. Transmittance Calculation (10 points)
1. 3D Gaussian Splatting
1.1 3D Gaussian Rasterization (35 points)
1.2 Training 3D Gaussian Representations (15 points)
lrs:
'lr': 5e-4, "name": "opacities",
'lr': 1e-3, "name": "scales",
'lr': 1e-3, "name": "colours",
'lr': 1e-4, "name": "means",
num_iters: 1000,
PSNR: 27.214; SSIM: 0.910;
final renders
training progress
1.3.1 Rendering Using Spherical Harmonics (10 Points)
GIFs
default
view dependent
Static Images
default
view dependent
default
view dependent
There are clearer shadows at the left side of the chair in view dependent images.
2. Diffusion-guided Optimization
2.3. Visualization
2.1 SDS Loss + Image Optimization (20 points)
a hamburger; w/o guidance 400 iters
a hamburger; w/ guidance 1999 iters
a standing corgi dog; w/o guidance 1999 iters
a standing corgi dog; w/ guidance 1999 iters
a teddy bear; w/o guidance 400 iters
a teddy bear; w/ guidance 900 iters
a classroom; w/o guidance 1999 iters
a classroom; w/ guidance 1999 iters
2.2 Texture Map Optimization for Mesh (15 points)
black color
front bright light
2.3 NeRF Optimization (15 points)
a standing corgi dog
a basketball
a teddy bear
2.4.1 View-dependent text embedding (10 points)
a standing corgi dog
a basketball
a teddy bear
The view dependent text condition brings a clearer rgb result, with more continuous depth. In the corgi case, view dependent one doesn't have multiheads artifacts.