Modulcode:	infCV3D-01a
Englische Bezeichnung:	3D Computer Vision
Modulverantwortliche(r):	Prof. Dr. Kevin Köser
Turnus:	unregelmäßig (SS22 SS23 SS24)
Präsenzzeiten:	4V 2Ü
ECTS:	8
Workload:	60 h lectures, 30 h exercises, 150 h self studies
Dauer:	ein Semester
Modulkategorien:	MSc-Inf-WP (MSc Inf (21)) 2F-MEd-Inf-WP (MEd-Hdl Inf (21)) 2F-MA-Inf-WP (2F-MA Inf (21)) MSc-WInf-WP-Inf (MSc WInf (21)) WI (MSc Inf (15))
Lehrsprache:	Englisch
Voraussetzungen:	Inf-Math-A Inf-Math-B Inf-Math-C

Kurzfassung:

This course is not only addressed to Master students, but also to advanced Bachelor students (Bachelor Wahlpflichtfach). The lecture will introduce the principles of 3D computer vision, i.e. how to make imaging tractable with equations and what concepts are needed to build 3D vision systems that are nowadays at the core of autonomous cars, augmented reality glasses or to create models for 3D printing: How to use the two views of a stereo system (like the human eyes), or a camera mounted on a moving robot, to infer the 3D geometry of the environments, and additionally the ego motion in a robust way. The key concepts are projective geometry, the pinhole camera model and multiple view relations, as well as robustly finding correspondences between different images of the same scene.

Lernziele:

The students understand the mathematical camera model and how cameras can be used to infer information about 3D scenes using the concept of correspondences. They can handle entities of projective geometry and image-based geometric relations and implement these in the context of 3D computer vision.

Lehrinhalte:

This lecture will introduce the basic principles of 3D computer vision, which are nowadays also used in robotics, augmented reality, smartphones, for terrain mapping such as for the deep ocean floor, or to create models for 3D printing, archaeology or computer games.

In order to treat machine vision in a principled way, the mathematical pinhole camera model is introduced, and students will learn the representations of points, lines and other entities in images and in space using projective geometry. Afterwards, multiple view relations will be discussed and how camera motion and scene geometry can be inferred from common observations in images, called correspondences. This involves linear algebra, optimization and robust estimation, since we have to deal with mismatches and outliers when automatically finding correspondences. Finally, we will look into how to integrate the different concepts into a complete approach for 3D vision from images, how to obtain dense surface models and see current applications and open problems in 3D vision.

The following topics are discussed:

• Image Processing, Corresponding Features
• Basics of Projective Geometry
• Pinhole Camera Model
• Homographies and Panoramic Mosaics
• Epipolar and Multiple-View Geometry
• Standard Stereo and Depth Estimation
• Camera Tracking and Pose Estimation
• 3D Vision System

Lectures, slides and course material will be in English.

Weitere Voraussetzungen:

Mathematical knowledge from Bachelor courses in linear algebra, geometry, analysis, and solving of linear equations is needed. Prior knowledge in image processing, like the Bachelor lecture InfEinfBV (Introduction to Image Processing) is required.

Prüfungsleistung:

Oral exam or written exam (depending on number of students), will be announced at beginning of the semester.

Lehr- und Lernmethoden:

This course will be taught as a lecture two times a week, complemented by exercises once a week.

For the exercise, weekly homework is handed out and solved in teams of 2. During the weekly exercise sessions, active participation of the students is expected. Each team should present homework solutions at least once to be accepted to the exams.

Verwendbarkeit:

Literatur:

Szeliski, Rick: Computer Vision: Algorithms and Applications. Springer 2010. Electronic preprint version available at: http://szeliski.org/Book/

Hartley, Zissermann: Multiple View Geometry in Computer Vision (2nd edition), Cambridge 2004.

Verweise:

Kommentar:

Due to large thematic overlap, students cannot combine the ECTS points of this module with points of the module "Image-based 3D Scene Reconstruction" of previous semesters.