A. Neural Volume Rendering
1.3. Ray sampling (5 points)
1.4. Point sampling (5 points)
1.5. Volume rendering (20 points)
2.1. Random ray sampling (5 points)
2.2. Loss and training (5 points)
Results
Box center: (0.25, 0.25, -0.00)
Box side lengths: (2.01, 1.50, 1.50)
2.3. Visualization
3. Optimizing a Neural Radiance Field (NeRF) (20 points)
4.1. View Dependence (10 bonus points)
View dependence allows rendering of specular surfaces that look different from different angles. Low view dependence will result in a more matte (diffuse) surface, and higher view dependence will allow for sharper glossy (specular) surfaces. Too much view dependence may result in the model “cheating” by generating colors to match just the angles of the input images and not generating a generalized output for other angles, decreasing generalization.
4.2. Coarse/Fine Sampling (10 bonus points)
Coarse: 64 stratified samples per ray. Fine: 128 additional samples from the coarse weights. 192 total.
Speed decreased due to the need to run sampling twice and increased total number of samples per ray total 128 -> 192. Quality goes up because we waste less samples on inner points that don’t contribute to the render and get more samples on useful points near the surface.
4.3. VolSDF (10 bonus points)
Alpha determines how dense the density of material is inside the object.
Beta determines how quickly the density falls off to zero outside the object (sharpness).
How does high
betabias your learned SDF? What about lowbeta?High
betaallows material to seep out infront of the surface more, resulting in a fuzzy, fatter object. Lowerbetacuts off material closer to the surface, creating a sharper object.Would an SDF be easier to train with volume rendering and low
betaor highbeta? Why?High
betais easier to train than lowbeta, The thinness of the surface region impacts how much of the sample provide gradients to the model while learning. Sharper surface region means less samples hit this region and provide gradients.Would you be more likely to learn an accurate surface with high
betaor lowbeta? Why?If both settings are guaranteed to converge, low
beta. Once trained, the sharpness and details can be maintained only by the sharper surface of the lowbeta.
| Setting | Value |
|---|---|
| alpha | 10.0 |
| beta | 0.05 |
| eikonal weight | 0.02 |
| epochs | 250 |
Ran a series of 10-epoch tests to empirically find good values.
I think beta follows the sample spacing which is around
0.03 apart. This allows samples to hit the surface region but keep it
minimal so allow sharp images. High alpha allows some
gradients to pass through so we can accumulate some samples and have a
solid surface, but not bleed out into the interior too much.
B. 3D Gaussian Splatting
5. 3D Gaussian Rasterization (35 points)
6. Training 3D Gaussian Representations (15 points)
| Parameter | Learning rate |
|---|---|
| opacities | 0.05 |
| scales | 0.01 |
| colours | 0.05 |
| means | 0.001 |
Iterations: 1000
PSNR: 29.168
SSIM: 0.935
7.1. Rendering Using Spherical Harmonics (10 points)
Brighter contrast of green and gold. Deeper shadows in the gold ornamentation.
Brighter contrast of green and gold. Darker and sharper shadows of the armrests and backrest.
Brighter contrast of green and gold. Darker and sharper shadows around ridges in the back and legs.
7.2. Training On a Harder Scene (10 points)
Baseline
Setup from 6: isotropic Gaussians, random initialization, L1 loss, 1000 iterations.
| Parameter | Learning rate |
|---|---|
| opacities | 0.05 |
| scales | 0.01 |
| colours | 0.05 |
| means | 0.001 |
PSNR: 18.866
SSIM: 0.685
Improved
| Baseline | Improved | |
|---|---|---|
| Gaussians | isotropic | anisotropic |
| Loss | L1 | 0.8 · L1 + 0.2 · (1 − SSIM) |
| Mean initialization | starter default | uniform in [−1.2, 1.2]³ |
| Initial scale / opacity | starter default | 0.02 / 0.1 |
| Iterations | 1000 | 3000 |
| PSNR | 18.866 | 23.654 |
| SSIM | 0.685 | 0.873 |
| Parameter | Learning rate |
|---|---|
| opacities | 0.05 |
| scales | 0.005 |
| colours | 0.025 |
| means | 0.002, decayed exponentially to 0.00002 |
| quaternions | 0.001 |
I’ve based off Claude’s smaller experiments exploring various hyperparams.
This run combines the modifications run on smaller sample runs into one improved attempt:
- 3000 iterations instead of 1000: more runs.
- Spread-out initialization: so Gaussians cover all the objects from the start.
- Low initial opacity, 0.1: Gaussians behind others still get gradients at start of training.
- Initial scale 0.02: Gaussians cover the scene without starting as fog.
- Anisotropic Gaussians: they can stretch and rotate to fit edges and highlights.
- L1 + SSIM loss (0.8 / 0.2): sharper structure.
- Means learning rate 0.002, decayed to 0.00002: fast moves early, stable settling later.
- Lower colour and scale learning rates: stable training with the extra per-Gaussian parameters.