AI and programming classes in Leith, Edinburgh

The learner takes a 498 m square from OpenStreetMap, lying wholly inside the area the map outlines as Leith. It holds 275 complete building outlines with 3,102 points between them. A point counts as a true corner when the outline turns there by 30 degrees or more, which gives 1,485 corners once points within a metre of each other are merged. The program then draws the buildings as white shapes on a black picture and forgets the outlines. The detector sees only pixels. For each pixel it measures how fast brightness changes across and down, combines the two into a Harris score, and keeps the local peaks above a chosen share of the strongest score. A detection is right if a true corner lies within 2 m.

Precision is the share of detections that are real corners; recall is the share of real corners that were found. At a 5% threshold the detector made 1,053 detections, 97.2% of them correct, and 82.0% of the true corners had a detection within 2 m. Drop the threshold to 1% and detections nearly double to 2,061, but recall only creeps up to 85.0% while precision collapses to 54.1%: almost all the extra detections are false. Raise it to 40% and nearly every detection is right, at the price of missing three corners in ten. Coarser pictures are worse on both counts. At 2 m per pixel the detector finds 16.4% of corners and fewer than half of its 457 detections are right; one pixel is then as wide as the whole 2 m matching tolerance. Recall can run ahead of the detection count because one detection may sit within 2 m of two close corners.

A Leith teenager who has scored a detector against known answers understands evaluation, the part of AI that matters most once models are easy to call. The longer argument is in why teenagers should still learn to code in 2026.