Learning 3D Reconstruction in Function Space

Explore a keynote presentation on learning neural implicit 3D representations for advanced 3D reconstruction. Delve into the departure from traditional explicit 3D shape representations using voxels, points, and meshes. Discover how implicit representations offer a small memory footprint and enable modeling of arbitrary 3D topologies at any resolution in continuous function space. Examine the capabilities and limitations of these approaches in reconstructing 3D geometry, textured 3D models, and motion. Learn about the development of implicit 3D models using only 2D supervision through an analytic closed-form solution for gradient updates. Cover topics including Occupancy Networks, Texture Fields, Occupancy Flow, and Differentiable Volumetric Rendering. Gain insights into the evolution of 3D reconstruction techniques from the 1963 Blocks World to modern neural implicit representations.

Intro
1963: Blocks World
Traditional 3D Reconstruction Pipeline
3D Representations
Occupancy Networks
Training Objective
Texture Fields
Representation Power
Occupancy Flow
Architecture
Differentiable Volumetric Rendering
Summary