Storey 04 / 05 / The Pagoda
Vision
Computer vision is teaching a computer to read images: find the things in a picture, follow them over time and measure them.
I train and run models in PyTorch. I have used them on football broadcasts, on radar images of forest and on gigapixel slides of tissue, where each slide is cut into tiles before a model can read it.
Much of the work sits around the model. The images have to be cleaned before they go in, and what comes out has to be checked against something real.
The tower
Tools 7
- PyTorchTraining the football event classifier and the deforestation model.Football Intelligence →
- ONNXRunning the football models at half precision, in batches, which brought the pipeline to 24 frames a second.Football Intelligence →
- YOLOv8Finding the players, goalkeepers, referees and ball in each frame of a match.Football Intelligence →
- ByteTrackLinking each player from one frame to the next, so each keeps one label.Football Intelligence →
- HoVer-NetFinding and outlining every cell nucleus in pathology slides, at the TIA Centre and in the society's teaching sessions.TIA Centre, University of Warwick →
- TIAToolboxCutting gigapixel slides into tiles, evening out their stain colours and filtering out poor tiles.TIA Centre, University of Warwick →
- Earth EngineFetching eight years of Sentinel-1 radar over 100,000 hectares of forest.Deforestation Detection →
Notes from this storey
A shaking map
In my football pipeline, the map from the camera view to the pitch is worked out again for every frame.
Slides too big to read at once
A pathology slide can hold billions of pixels, far too many for a model at once. At the TIA Centre I cut each slide into tiles and filtered out the poor ones.
Standing on this storey
Project / 2026
Football Intelligence
You upload a video of a football match, filmed by the ordinary television camera.
The right label stays on the right player about 92% of the time.
Read the chapterProject / 2026
Deforestation Detection
A system that finds where forest has been cut down, even when clouds hide it from ordinary satellite photos.
About 65 square kilometres of cleared forest, roughly 9,000 football pitches.
Read the chapter