Football analysis with computer vision

What it does

It takes a match video and extracts three things: where every player is, which team they belong to, and how much and how they moved. On top of that it builds position heat maps and distance and speed statistics.

What actually has merit, and what does not

The interesting part of the project is the homography. A football camera does not look at the pitch from above or from a constant angle: it looks from a stand. To go from pixels to metres you have to estimate the transformation between the image plane and the pitch plane, and for that you need recognisable keypoints.

I trained a detector for those keypoints because generic models cannot distinguish the lines of a football pitch from any other lines. With the homography solved, distance covered and speed become metres and metres per second instead of pixels.

What has no merit is presenting it as a system. They are notebooks: the detection is solved, the engineering around it does not exist, and there is not a single metric I can show. A model that cannot be retrained because the weights are not in the repository is not reproducible, it is a screenshot with code.

What I learned

That in computer vision 80 % of the work is preparing the data and 20 % is the model, and that this split never shows in tutorials because tutorials start with the dataset already prepared.

And that a portfolio project without numbers does not convince anyone, however good the visual result. This one is in my portfolio precisely for that reason: it shows what I know how to do and shows, with no dressing up, how far I got.