< back_to_lab
OBJECT DETECTION · FIELD REPORT

Site Safety Detection

A fine-tuned YOLO detector that looks at a construction scene and picks out the people in it along with the safety equipment they are — or are not — wearing. Trained today; this page is a placeholder for a proper write-up once the results have settled.

object detectionYOLOfine-tuningcomputer vision

01 · THE MODEL, RUNNING

Three scenes, one detector

Real photographs the model had not seen, run through the detector and drawn with its own boxes and confidences. The last scene is the interesting one: a person on a rooftop slab with no protective equipment on at all.

demo-reel.mp4 · fine-tuned detector
02 · WHAT IT LOOKS FOR

Four classes

The detector was fine-tuned to find one subject and three pieces of equipment, so that “is this person wearing what they should be?” becomes a question about which boxes overlap.

Person

The anchor class. Everything else is judged relative to the people actually in frame.

Helmet

Hard hats, including at distance and at awkward angles where only part of the shell is visible.

Hi-vis vest

High-visibility clothing, which is easy in bright conditions and genuinely hard in shade and glare.

Gloves

The smallest and least forgiving of the four — often partly occluded by tools or the work itself.

03 · STATUS

Where it stands

1
TRAINED, AND IT DETECTS

The fine-tune is done and the model finds all four classes on held-out photographs, including the negative case where a person is present and no equipment is.

2
NOT YET MEASURED IN PUBLIC

No precision, recall or mAP numbers on this page yet — partly because a single day of training does not earn them, and partly because the failure cases are more interesting than the headline metric. Both are coming.

3
NEXT

Video rather than stills, a proper evaluation split, and the honest account of what the detector misses — small gloves, heavy occlusion, and hi-vis in direct glare.