Course Overview

A three-week course on how computers interpret images, from the physics of image formation and classical image processing to the deep networks used today for classification, detection and segmentation, and on to generative models, vision–language models and responsible deployment. The course is built around intuition and real examples: results in the lectures come from code run for the course on real photographs and public datasets.

Course website: slides, notebooks, assignments and project brief

Course Structure

Fifteen two-hour lectures, one per weekday, with tutorials and labs run separately by the tutors.

• Week 1, Foundations: human vision; image formation, sampling, quantisation and colour; noise, convolution, filtering and histograms; edges, corners, SIFT and HOG; from engineered to learned features.
• Week 2, The deep learning era: CNNs and how to train them; architectures from AlexNet to ConvNeXt and transfer learning; object detection from R-CNN to YOLO; semantic and instance segmentation and Segment Anything; Vision Transformers, text-guided localisation and image captioning.
• Week 3, Generative vision and the road ahead: autoencoders and VAEs; diffusion models; CLIP and vision–language models with applications across domains; robustness, fairness and responsible deployment; capstone presentations.

Assessment

• Weekly quizzes (60 minutes)
• Individual LaTeX-typed assignments in Weeks 1 and 2
• A team capstone project on a real-world computer vision problem, presented on the final day

Prerequisites

• Python programming
• Linear algebra, calculus and basic probability