16-825 Assignment 3

1.3

1.4

1.5

2.2

Box center: (0.25, 0.25, -0.00)

Box side lengths: (2.01, 1.50, 1.51)

2.3

3

4.3

alpha is the density inside the object. beta controls how smooth the density changes around the surface: density goes from alpha inside to 0 outside within a band of width about beta around the zero level set.

1. High beta makes the density blurry, so the learned SDF is smooth and the surface can be thick or offset. Low beta makes density close to a step function, so the SDF gives a sharp surface.

2. High beta is easier to train. Density is non-zero over a wider range, so many samples along a ray get gradients and the optimization is more stable. With low beta only points very close to the surface get gradients.

3. Low beta. Density is concentrated at the zero level set, so the rendered color comes from the surface itself and the surface is sharp and accurate.

Settings: alpha = 10, beta = 0.05, eikonal weight = 0.02, 6-layer 128-unit distance MLP with softplus, 2-layer color MLP, 6 harmonic functions, lr 5e-4, 250 epochs. beta = 0.05 is small enough for a sharp surface and big enough for stable training. The eikonal loss keeps the network a valid SDF.

5

6

Learning rates: opacities 0.05, scales 0.005, colours 0.02, means 0.0005

Iterations: 1000

PSNR: 29.578

SSIM: 0.936


7.1

Without SH (5.5):

With SH (7.1):

Left: without SH, right: with SH.

With SH the back of the chair has stronger contrast and highlights in the pattern. The golden parts look shiny and the seat is darker on one side, because the color depends on view direction.

With SH the green cushion and the frame have more shading and darker regions. Without SH the colors look flat.

Small difference. The bottom of the seat is a diffuse surface, so the view-dependent terms are small. With SH the pattern on the cushion is slightly brighter.