Computer Vision
From Filters to Foundation Models
Computer Vision
COMS4036A & COMS7050A
Semester 2, 2026
School of Computer Science and Applied Mathematics
Preface
This book accompanies the Honours/MSc Computer Vision course at the School of Computer Science and Applied Mathematics, University of the Witwatersrand. It follows the course week by week. There are fourteen teaching weeks, each with a long-form lecture and, in all but the last, a two-hour lab. It is also written to be readable on its own, front to back.
The course outline sets out the administration: timetable, syllabus, prerequisites, grading, and policies. The schedule gives the week-by-week plan.
How the book is organised
The course teaches detection and segmentation twice: first with the tools of the CNN era (Part II), then again after transformers arrive (the end of Part III). This is deliberate. You will solve the same problems with two different toolkits and then compare them: anchors and non-maximum suppression against learned queries and set prediction. It is also, roughly, how the field itself experienced the years 2020–2023.
Two chapters carry a self-study badge: Cameras and Multi-View Geometry (Chapter 16) and Video Understanding (Chapter 18). They are not lectured, but they are positioned where the logic of the subject puts them, not in an appendix. If your project touches 3D reconstruction, read the geometry chapter before the neural rendering one. Chapter 17 assumes you know what a camera matrix is.
The threads
Two threads run through every chapter, marked by coloured badges in the text.
The skills ladder. Each lab introduces one piece of practitioner craft alongside its content: experiment tracking with Weights & Biases, the overfit-a-single-batch debugging ritual, writing a Dataset class, Bayesian hyperparameter sweeps, mixed precision, reading a model card. These are collected as a reference handbook in Appendix A — The Practitioner’s Toolkit. The ladder is examinable. Test 2 is the practical test, and it draws on the ladder directly.
Ethics and bias. Rather than a standalone chapter that nobody connects to the technical material, each chapter raises one concrete question at the moment it becomes technically natural: where benchmark images come from, what transfer learning inherits, whose faces detection fails on, what a generative model’s training set licenses it to do. These are short, specific, and tied to the method on the page.
Labs
Every chapter ends with a lab pointer; where a chapter has no lab of its own, the pointer says so. The labs assume a free-tier Colab GPU (a T4 with roughly 15 GB of memory); anything heavier uses pretrained checkpoints. Appendix C — Lab Index lists the thirteen labs with the chapters they pair with and the rung of the skills ladder each one adds.
Prerequisites
Multilayer perceptrons, backpropagation, and basic optimisation are assumed. The CNN chapter starts from “you know what a learned weight is”. Probability at the level needed for diffusion models, and the linear algebra of attention, are refreshed in Appendix B — Math Refresher.