Semester 3 · Computer vision
Image Formation and Classical Computer Vision
https://epochschool.com/subjects/image-formation-classical-vision
How cameras turn the 3-D world into arrays of numbers, and how to get geometry and meaning back out of them without a neural network.
You derive the camera model, filtering and the frequency domain, edge, corner and blob detectors, descriptors and matching, homographies and RANSAC, calibration, two-view geometry and stereo, optical flow and classical segmentation, and you implement every one of them in NumPy and SciPy. Five CPU labs grade your detectors, stitcher, calibrator, stereo matcher and flow estimator against exact ground truth. These are the foundations that the deep vision, generative and 3-D subjects of the CV track build on.
- Prerequisites
- Required: Finish this subject before you start. The subject builds directly on it.Recommended: Helpful background. You can start without it if you are willing to fill the gaps.
What you’ll be able to do
- LO1Derive and apply the pinhole camera model x = K[R|t]X in homogeneous coordinates, including intrinsics, extrinsics, vanishing points and lens distortion, and predict how image measurements change with depth and camera parameters.
- LO2Explain how light, sensors and colour encoding produce pixel values, and convert between linear and gamma-encoded light, colour spaces and point operations correctly on NumPy image arrays.
- LO3Design, implement and analyse linear and non-linear filters in the spatial and frequency domains, including convolution and frequency response, separability, the DFT and DCT, sampling and aliasing, interpolation and image pyramids.
- LO4Implement and justify detectors for edges (Canny), lines (Hough), corners (Harris, Shi–Tomasi) and blobs across scale (LoG, DoG), and analyse their invariance and repeatability.
- LO5Build gradient-histogram descriptors (HOG and a SIFT-style descriptor), match them with ratio and mutual tests, and evaluate matching quality against ground truth.
- LO6Estimate 2-D transformations and homographies with the normalised DLT and RANSAC, and use them to warp and stitch images, diagnosing when a homography model fails.
- LO7Calibrate a camera (Zhang's method with non-linear refinement), estimate two-view geometry (F, E, relative pose, triangulation) and compute depth from rectified stereo, quantifying the error in each.
- LO8Estimate optical flow with Lucas–Kanade (including coarse-to-fine) and Horn–Schunck, explain the aperture problem and the failure modes, and evaluate flow with endpoint error.
- LO9Segment images with thresholding, morphology, clustering and graph-based methods, evaluate the results with IoU, Dice and Rand-type indices, and judge when classical methods beat learned ones.
Chapters
The chapter and lab list for this subject is being finalised and will be published here.
Before you start
Required
Finish this subject before you start. The subject builds directly on it.
Neural Networks & Deep Learning
Recommended
Helpful background. You can start without it if you are willing to fill the gaps.
Foundations of Machine Learning
Probability & Statistics for ML
Sequence Models & Transformers
Deep Visual Recognition
Deep Generative Models
Vision-Language, 3D Vision and World Models
Capstone Project
How it’s taught
You read short sections written for the page. Each ends with a check, and in-browser exercises let you practise. Graded labs run on your own cloud lab machine, CPU-friendly.
No videos. No live classes. No AI tutor.