Andrea Tagliasacchi: Structured Representations for 3D Computer Vision

Explore cutting-edge techniques in 3D computer vision through this 49-minute talk by Andrea Tagliasacchi from Google Brain Toronto. Delve into three key areas of 3D scene understanding: permutation equivariant learning for robust optimization, pose-conditioned implicit functions for digital human representation, and a hybrid implicit/explicit differentiable representation of 3D geometry. Learn about the Attentive Context Network (ACNe), Neural Articulated Shape Approximation (NASA), and innovative approaches to 3D representation that combine the training ease of implicit functions with the efficiency of polygonal meshes at inference time. Gain insights into the latest advancements in 3D sensing, capture, tracking, compression, modeling, and simulation of geometry from an expert in the field.

Intro
3D scene understanding
Permutation equivariant leaming
Attentive Context Network (ACNe)
Attentive Residual Block (ARB)
Attentive Context Normalization (ACN)
Summary - acne.github.io
Traditional digital humans
Neural digital humans - NASA
Rigid model (R)
Reconstruction on AMASS
CP» free ICP registration
Summary - NASA
What is the best» 3D representation?
Convexes: why are they relevant?
Universal approximator of convex domains
Implicit functions @ training time
Polygonal meshes @ inference time
Related work @ CVPR 2020