Definition
An object detection model is a program trained to indicate which objects appear in an image and where. The output is a list of boxes with an assigned class and confidence value, for example person, helmet, forklift.
How the model knows what to look for
The model learns from a set of images in which a human marked the objects of interest. The closer those images are to the target conditions, the better the model performs after deployment.
That is why a model trained on construction site photographs can be worse in a production hall, even though both look for helmets. Lighting, camera angle, background and typical clothing all differ.
Customer-specific fine-tuning
Fine-tuning means further training on material from a specific plant. It is used when something unusual occurs: an atypical colour of workwear, specialised equipment, a rare type of violation.
It requires collecting and labelling material, so it makes sense where an off-the-shelf model genuinely fails, not as a default part of every deployment.
Frequently asked questions
How many images does fine-tuning need?
It depends on the task and how far it differs from the baseline data. Usually hundreds of examples per class rather than thousands.
Does our footage leave the site?
That is a contractual matter. Fine-tuning can be done on material shared within an agreed scope, or locally.
Do models age?
Changes on site - new signage, different clothing, a rebuilt hall - affect accuracy. That is why reviews also cover detection quality.