Homework 4 Tianyuan

andrew id: tianyuaz

late day: 1

Problem 1

Problem 2

I useed regularization term to be 0.04

Problem 3

1. Low beta make the density distribution more smooth, more volume rendering. Low beta makes the volume distribution more sparse, more sharp, make shapes better, but probably harder to converge
2. I think it is easier to train with hight beta, because the distribution get more smoother, and more well behaved gradients.
3. But for accurate shapes, I would prefer lower beta, which produce sharper distributions, and makes the final shape more accurate

Problem 4

I chose the neus.

I implemeneted neus sdf to density convertion function. use the scaling of 3.0. The shape of the same scene as in problem-3 is below:

Seems harder to converge