Course Title

Dynamic Vision: Tracking Methods in Computer Vision


Course Description

Tracking has been a widely studied task in Computer Vision with many different applications and a large variety of task formulations. The task can be formulated from short-range but dense motion estimation like optical-flow to long-range tracking of single objects. We take a look at the different tasks and present the state of the art in each of them. We review optical flow, point-, feature and object tracking. We furthermore present the latest trends in test-time optimization-based methods, for example dynamic gaussian splatting.


Course Organizer

Friedhelm Hamann


Course Format

Hybrid seminar. The idea is to have interchanging lectures given by the PhD student and by an invited speaker. In the first 1-2 lectures of a block a task will be introduced and in the last lecture an invited speaker will introduce the latest influential work in the field.


Target Group

CS/EE/IT students


Course Structure

  1. Overview
  2. Optical flow: Intro/non-learned (Concept, KLT)
  3. Optical flow: Current SOTA, learned (RAFT, Tracking every pixel)
  4. Point tracking: Intro/non-learned (Particle Video)
  5. Point tracking: Current SOTA, DL-methods (Point Odyssee)
  6. Feature tracking: Intro/non-learned
  7. Feature tracking: Current SOTA, DL-methods, Applications like VO
  8. Object tracking: Intro/non-learned
  9. Object tracking: Current SOTA, DL-methods
  10. Multi-object tracking: Intro/non-learned (SORT)
  11. Multi-object tracking: Current SOTA, DL-methods (TAO)
  12. Dynamic NeRF, Dynamic Gaussian Splatting
  13. Summary