ACIT4030 Machine Learning for images and 3D data Course description

This course will present the state of the art in algorithms for machine learning on images and 3D data. After a brief introduction to image processing and 3D geometry, we will cover topics within both supervised and unsupervised learning. The course covers classical problems like classification, segmentation, and correspondence detection. Recent work on shape and image synthesis will also be discussed. We will in particular study deep neural architectures for 2D images and 3D data such as point clouds and shape graphs. Additionally, 3D shape design with generative models will be presented.

We assume good background in programming, machine learning, and linear algebra. Knowledge of computer graphics and image processing is preferable, but not strictly required.

No formal requirements over and above the admission requirements.

Upon successful completion of the course, the candidate:

Knowledge

• has knowledge of classical problems within graphics and imaging that are applicable to machine learning, such as classification, segmentation, and correspondence detection.
• has a good understanding of problems related to generation of new images and 3D shapes.

Skills

• is able to apply state-of-the art machine learning algorithms to real-world problems related to imaging and 3D graphics.

Competence

• is aware of the state of the art in algorithms for machine learning on images and 3D data.
• has experience with real world problems within the course domain, with a focus on solutions using deep neural architectures.

• Introduction to image processing and geometric modelling
• Convolutional neural networks for images and graphs
• Segmentation for images and shapes
• Correspondences and mappings
• Modelling, synthesis, and analysis
• Joint embedding for images and 3D data

Teaching approach is a combination of traditional weekly lectures and assignments, student- led seminars, and a final project. Lectures will present the core theory of the course content and homework will focus on theoretical knowledge. In student-led seminars, topical research papers will be presented and discussed. The final project exposes the student to a chosen real- world problem relevant to the course topic.

Practical training

The student will be exposed to programming with repositories such as ImageNet and ShapeNet and will have created solutions for real-world problems related to data-driven graphics and imaging.

The following required coursework must be approved before the student can take the exam:

Two mandatory group assignments consisting of technical tasks, summarized in reports (about 10 pages each).

• Individual student presentation (20 %)

• One individually written evaluation of another student presentation (1-2 pages) (10%)

• Individual final project report (between 25 and 35 pages) (70 %) All exams must be passed in order to pass the course. The assessment of the presentation cannot be appealed.

All printed and written aids and a calculator that cannot be used to communicate with others.

For the final assessment a grading scale from A to E is used, where A denotes the highest and E the lowest pass grade, and F denotes a fail.

Two internal examiners. External examiner is used periodically.