Object detection

Detection underlies counting, PPE compliance, intrusion alerts and defect localisation. YOLO-family architectures are the practical default for real-time work.

Small, overlapping or densely packed objects remain genuinely hard and usually need a different approach.

Class imbalance is the usual practical problem. Real production lines produce thousands of good units for every defective one, so a detector trained on raw collection will be excellent at finding nothing. Deliberate oversampling of the rare class and per-class metrics are what make the result honest.

Annotation quality sets the ceiling. Two labellers who disagree about what counts as a defect will produce a dataset with contradictory examples, and no amount of training compensates for that. Agreeing the definition, with photographs, not adjectives, is the first task, not the last.

Report per-class metrics rather than an aggregate. A detector that is excellent on the common defect and poor on the rare one can post a flattering overall number while missing exactly the cases the system exists to catch.