Previous | Next --- Slide 13 of 28
Back to Lecture Thumbnails
sitongc

How to caculate the gradient? Is there any formula for this?

Senbei

This process reminds me of the gradient descent.

I guess for doing this, we should carefully choose the step size or it will never converge. And we also might need a proper threshod for judging wheather it is converged or not.

EmDeeZee

What is the convergence threshold? If it's too high, then you're not really converging to a point on the surface (instead converging to a point just off of it). If it's too low, it may take a lot longer to converge. How is this threshold chosen?

Zhuoqian

People are doing some research on a 3D data structure called "signed distance functions" which uses this idea. I first read about this in "DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation". https://arxiv.org/abs/1901.05103