3D Reconstruction with Differentiable Rendering

Course Description

This course will focus on generic 3D reconstruction of objects from multiple images. We will start with an introduction to the classical geometrical approach to this problem. Then, we will study and derive the differentiable approach for rendering. Thanks to the differentiability, the whole pipeline can be built in PyTorch with standard optimization frameworks such as ADAM to minimize reprojection error. The rest of the course will explore different applications of differentiable 3D reconstruction such as object tracking, deblurring, and 3D modeling.

Course tools

• Python, NumPy, PyTorch, PyTorch3D/Kaolin

Prerequisites

• Linear Algebra
• Computer Graphics
• Projective Geometry
• Basics of PyTorch

Level of complexity of course Intermediate

Lecturer

Mr. Denys Rozumnyi ETH Zurich, Ph.D. Student Denys’s main expertise is in deblurring and 3D reconstruction of moving objects. During his undergraduate studies, Denys focused on the detection and deblurring of simple fast-moving objects. More recently, as part of my Ph.D., he has been focusing on reconstructing 3D models of objects with complex shapes that are significantly blurred. Currently, Denys is also interested in the general 3D reconstruction of scenes and objects. Fields of interests: 3D reconstruction, deblurring, deep learning Contacts:denys.rozumnyi@inf.ethz.ch https://www.linkedin.com/in/denys-rozumnyi-35004a145/ https://twitter.com/DRozumnyi https://www.facebook.com/rozumden/