Assignment 3

Box Center (rounded to nearest 1/100): (0.25, 0.25, 0.00) Box Side Lengths (rounded to nearest 1/100): (2.01, 1.50, 1.50)

View dependence can capture specular details to show them with better realism, but risks overfitting to sparse training views, causing artifacts at novel viewpoints.

Sphere tracing marches along rays by the SDF value at each point, iterating until the distance falls below a threshold or maximum steps are reached.

The MLP: fully connected layers with ReLU, with no final activation since SDF values can be negative. The eikonal loss enforces ||d f(x)|| = 1, for a valid distance function.

I used moderate beta (0.1) with increased eikonal weight to balance training stability and surface accuracy, providing strong gradients while maintaining geometric detail.

The NeuS naive approach uses density proportional to the SDF gradient's Laplacian, providing noisier gradients. VolSDF's sigmoid-based transformation offers smoother optimization.