Venue: Skolkovo Institute of Science and Technology
Taught for: Master of Science in Data Science (other programs eligible)
Abstract
Geometry plays an extremely important role in many computer vision algorithms as certain kinds of geometric transformations (e.g., projective) form the basis of imaging, estimation, and reconstruction. This course focuses on processing the geometry of 3D scenes and shapes, as obtained from both images and depth sensory data, using a series of learnable approaches. We will cover the standard geometry processing pipeline, study the depth acquisition systems, and dive into a variety of deep learning methods defined on semi-structured and unstructured geometric datatypes. Geometric learning-based systems differ from conventional ones by needing a custom way to construct low-level building blocks such as convolutional operations, that do not naturally exist for many geometric data structures. To this end, we will consider both familiar structures such as 2D images and 3D volumetric grids, and purely geometric ones such as point sets, meshes, implicit functions, and CAD representations such as parametric models. The course extensively leverages python programming skills focusing on numerical libraries such as numpy/scipy/pytorch, and requires basic knowledge of deep learning. Most of the software used within the course will be provided as docker images, thus knowledge of C++ or other tools should not be required.
Course Contents
Lecture 1. Course introduction and overview. The geometry processing pipeline. 3D representations in vision and graphics
Lecture 2. Hardware systems for 3D data acquisition
Lecture 3. Dense [2D] range images
Lecture 4. Point set-based modalities. Invariance and equivariance in learning.
