Machine Learning for Computer Vision

Machine Learning methods are an essential component for the solution of important problems in computer vision, including object classification and pose estimation, object tracking, image segmentation, denoising of images, or camera calibration. Therefore, in this lecture the most relevant methods of Machine Learning are presented and derived mathematically. These mainly comprise:
- kernel methods, specifically Gaussian processes
- metric learning
- clustering such as GMMs or spectral clustering
- boosting and bagging
- hidden Markov models
- neural networks and deep learning*
- sampling methods, specifically MCMC

The focus here is laid on a broad understanding of these methods rather than in a deep specification of single approaches. Practical experience is acquired by means of programming tasks.

*The topic “deep learning” will be handled only marginally. For a broader treatment of this topic, we refer to other classes, e.g. IN2346.

After successful participation in this module, the students dominate the basics of the most relevant machine learning methods for the field of Computer Vision. They are then able to give the underlying mathematical formulation of methods like boosting, bagging, HMMs, Gaussian processes, or MCMC, and they can associate these methods with an adequate application context in the field of computer vision. Furthermore, the are able to develop simple implementations of these methods and to apply them to concrete data sets.

Basic knowledge in linear algebra, calculus, and probability theory.
Statistical modeling and machine learning (IN2332)

In the lecture, there will be slides presented, and important mathematical formulations will be derived on the board. In the accompanying tutorials, practical and theoretical problems will be handled. These problems will be provided for home work sufficiently in advance to the dates of the tutoriels.

Slide show, blackboard

Christopher Bishop: Pattern Recognition and Machine Learning
Kevin Murphy: Machine Learning: A Probabilistic Perspective
Carl Edward Rasmussen and Christopher Williams: Gaussian
Processes for Machine Learning

The assessment is done in form of a 90 minute write exam. In this written exam, the students should prove that they understand the basic concepts of various different Machine Learning methods, and that they are able to apply them to specific problems in computer vision. The comprehension of the different methods mainly comprises the ability to describe and to derive them mathematically, but also the ability to discuss their benefits and drawbacks. Furthermore, application oriented problems are given to assess the students’ ability to apply adequate machine learning methods to specific tasks in computer vision, and also to group them accordingly (e.g . wrt. classification, regression, MLE, MAP, etc.).