Explore a comprehensive talk on physically-motivated learning for shape, material, and lighting recovery in complex scenes. Delve into the challenges of inverse rendering and image formation, examining novel approaches that combine deep learning with physical insights. Learn about differentiable rendering layers, parsimonious representations, and synthetic datasets that enable applications in augmented reality and scene editing. Discover how these techniques can transform mobile phones into accessible AR devices, and gain insights into spatially-varying lighting estimation, object insertion, and material editing using single unconstrained images.
Physically-Motivated Learning of Shape, Material and Lighting in Complex Scenes
Overview
Syllabus
Intro
Image Formation is a Complex Process
Component Problems of Inverse Rendering
Ambiguities of Inverse Rendering
Approaches to Inverse Rendering
A Canonical Challenge in Inverse Rendering
Outline
Image Formation: Rendering Equation
Background: BRDF
Background: Lighting
Large-Scale Dataset of Complex Materials
Physically-Based Rendering Layer
An Example
Comparison with Other Methods
Generalization to Real Data
Single Image Inverse Rendering
Physically Motivated Network: Rendering Layer
Synthetic Experiment: Global Illumination
Physically Motivated Network: Cascade Structure
Example: Cascade Structure
Example: Shape and Material Estimation
Example: View Synthesis
High-Quality, Photorealistic Augmented Reality
Key New Challenge: Spatially-Varying Lighting
Lighting Estimation Methods
Inverse Rendering in Indoor Scenes: Challenges
Ground Truth for Inverse Rendering Is Non-Trivial
Comparisons of Rendered Images
Compact and Effective Physical Lighting Representation
Spatially Varying Lighting Estimation: Representation
Physically-Motivated Network for Indoor Scenes
Spatially Varying Lighting Estimation: Results
Inverse Rendering in Real Indoor Scenes
Inverse Rendering: Quantitative Results
Object Insertion with Single Unconstrained Image
Object Insertion: User Studies
Material Editing with Single Unconstrained Image
Lightweight Acquisition with a Mobile Phone Camera
Physically-Motivated Deep Network
Conclusions
Taught by
Andreas Geiger
