Geometric structures and why (and how) to use them in deep models?

Abstract

In the last decades, the computer graphics community has invented many geometric structures to serve as useful representations to facilitate geometric modeling. As deep learning is transcending the 2D image space into the 3D realm through more efficient architectures, it may seem appealing to obsolete many previously developed geometric structures in exchange for more flexibility and generalizability. In this talk, I will convince you with several counter-examples why traditional geometric structures are valuable assets in the era of deep learning and how to leverage them to address typical shortcomings of neural networks in the context of geometry modeling.

Yifan Wang

(also publish as Wang Yifan)

Research Scientist at Adobe