Assignment 3 : Neural Volume Rendering and Surface Rendering

1. Differentiable Volume Rendering

1.1.5 Perform Splatting

1.2 Training 3D Gaussian Representations

Number of iterations: 300

opacities: lr- 0.01

scales: lr - 0.005

colours: lr - 0.01

means: lr - 0.001

Mean PSNR: 29.739

Mean SSIM: 0.943

training_progress

training_final_renders

1.3.1 Rendering Using Spherical Harmonics

Without Spherical Harmonics

With Spherical Harmonics

Comparison: With SH, the lighting and materials appear more realistic and with richer texture. The shadows are also softer.

2. Optimizing a basic implicit volume

2.1 SDS Loss + Image Optimization

a hanburger

a staning corgi dog

a car

a tree

2.2 Texture Map Optimization for Mesh

a dotted black and white cow & a pink cow

2.3 NeRF Optimization

python Q23_nerf_optimization.py --prompt "a standing corgi dog" --lambda_entropy 1e-4 --lambda_orient 1e-2 --latent_iter_ratio 0.2

python Q23_nerf_optimization.py --prompt "a color bird" --lambda_entropy 1e-4 --lambda_orient 1e-2 --latent_iter_ratio 0.2

python Q23_nerf_optimization.py --prompt "a tree" --lambda_entropy 1e-4 --lambda_orient 1e-2 --latent_iter_ratio 0.2