Project 3

bojiang

A. Neural Volume Rendering

1.3. Ray sampling

Pixel coordinates
Pixel coordinates
World-space ray directions
World-space ray directions

1.4. Point sampling

Points sampled along rays from the first camera
Points sampled along rays from the first camera

1.5. Volume rendering

Rendered box
Rendered box
Depth divided by its maximum value
Depth divided by its maximum value

2. Optimizing a basic implicit volume

After 1,000 epochs, the box center is (0.25, 0.25, 0.00) and its side lengths are (2.01, 1.50, 1.50), rounded to 0.01. The optimized silhouette, position, and rainbow appearance closely match the provided reference.

Optimized box
Optimized box

3. Optimizing a NeRF

The position-only Lego NeRF was trained for 250 epochs at 128×128 resolution. Across 200 test views: PSNR 24.32 dB, SSIM 0.8871. The main shape is reconstructed, while fine details remain soft.

Lego NeRF
Lego NeRF

4.1. View dependence — selected extension

Color depends on position and viewing direction; density depends only on position. A small color head with two direction-encoding frequencies limits angular capacity. View dependence can represent reflections, but greater angular capacity can memorize training cameras and generalize poorly between views. Fewer direction frequencies or angular regularization can improve generalization while smoothing sharp highlights. This Materials reconstruction retains blurred reflections.

View-dependent Materials NeRF
View-dependent Materials NeRF

B. 3D Gaussian Splatting

5. Gaussian rasterization

DC-only chair: RGB, depth, silhouette
DC-only chair: RGB, depth, silhouette

6. Training 3D Gaussian representations

Trained 10,000 isotropic Gaussians for 1,000 iterations with Adam and L1 RGB loss. Across all five held-out views: PSNR 29.19 dB; SSIM 0.9366.

ParameterLearning rate
Means0.0001
Log-scales0.005
Opacity logits0.05
Colors0.01
Training progress: predictions above ground truth
Training progress: predictions above ground truth
Final truck renders
Final truck renders

7.1. Spherical harmonics — selected extension

The pretrained chair is rendered with degree-3 spherical-harmonic color. The DC-only GIF is shown in Section 5; the full-SH GIF is below. Geometry and cameras are identical in both modes.

Full-SH chair: RGB, depth, silhouette
Full-SH chair: RGB, depth, silhouette
Frame 0: DC-only left, SH right. SH darkens parts of the green chair back and adds variation to its shading.
Frame 0: DC-only left, SH right. SH darkens parts of the green chair back and adds variation to its shading.
Frame 12: DC-only left, SH right. SH darkens the seat and lower back, increasing contrast in the gold fabric pattern and carved trim.
Frame 12: DC-only left, SH right. SH darkens the seat and lower back, increasing contrast in the gold fabric pattern and carved trim.
Frame 24: DC-only left, SH right. Brightness changes are smaller in this underside view. Both retain stretched Gaussian streaks; changing color does not correct geometric artifacts.
Frame 24: DC-only left, SH right. Brightness changes are smaller in this underside view. Both retain stretched Gaussian streaks; changing color does not correct geometric artifacts.