Tutorials

Vukašin Stanojević. Multiple Object Tracking – Challenges and Proposed Solutions

Vladimir Stanovov. Automated algorithms design – from genetic programming to large language models

Multiple Object Tracking – Challenges and Proposed Solutions

Multiple Object Tracking (MOT) remains an open and challenging problem in video understanding and computer vision. The most widely adopted solution is the tracking-by-detection (TbD) paradigm, which separates the task into two main steps: detecting objects in each frame and associating them with the currently tracked objects.

In this tutorial, we will address some of the unresolved subproblems in TbD-based MOT algorithms. The focus will be on handling unreliable detections, designing similarity measures for associating detections with tracked objects, and improving the tracking module (Kalman filter). We will review standard approaches to these problems, their advantages and drawbacks, and present some novel methods for improving the performance of multiple object tracking algorithms.

Automated algorithms design – from genetic programming to large language models

The challenge to automate computer programming has been around since the first computers. There were many attempts to approach this problem, one of which is genetic programming (GP). While the simplest GP systems, like tree-based GP, allowed creating small snippets of code, more advanced variants, like PushGP relied on a specifically designed programming languages.

Recently the success of large language models (LLMs) has led to the explosive growth of automated code generation, although LLM’s problem solving skills remain to be questionable. Still, many companies and researchers use LLMs as coding agents, as they allow speeding up the process. Modern LLM-based tools for automated algorithm design rely on evolutionary concept.

In this tutorial these two different approaches to automated algorithm design, starting from simplest GP systems to modern PushGP variants, as well as LLM-based tools, such as FunSearch and evolution of heuristics (EoH) will be presented. We shall discuss the main advantages and issues of such approaches, see what they are capable of and how to get started with them.