How to train your (neural) dragon

SIBGRAPI 2023 Tutorials - November 6, 2023

Neural fields have emerged as a promising framework for representing different types of signals. This tutorial focus on the existing literature and shares practical insights derived from hands-on experimentation with neural fields, specifically inapproximating implicit functions of surfaces. Our emphasis lies in strategies leveraging differential geometry concepts to enhance training outcomes and showcase applications within this domain.

The tutorial will be a half-day event (9:00 am to 12:00pm). 

In this tutorial, we will explore groundbreaking strategies, taking into account state-of-the-art shape representations using Neural Networks (NFs). We will present practical approaches to model and train NFs to approximate Signed Distance Functions (SDFs). We will also introduce methods for improving the training performance of NFs by implementing a strategy to dynamically sample data points during training. Finally, we will leverage the fact that an NF is a smooth function to analytically estimate the curvature measures of its level sets (no discretization is required).



Click to expand or download

How_to_train_your_neural_dragon (3).pdf


Coming soon