Object Detection — Level 4: YOLOv8 on a Manufacturing Scene

Approach

Same YOLOv8n model, pointed at a real warehouse floor. The pivot point of the whole project.

Workers in hard hats and safety vests walking an aisle of industrial equipment — a forklift, boxes, electrical gear all clearly visible in frame.

Original, unannotated warehouse aisle photo with a forklift, boxes, and workers — **Before** — the original, unannotated photo.

YOLOv8 on the same warehouse photo, only three person boxes drawn, forklift and equipment left unlabeled — **After** — `person: 3` and nothing else. Zero forklifts, pallets, boxes, or electrical equipment detected, despite all being clearly visible.

Terminal output

image 1/1 warehouse.jpg: 640x640 3 persons, 250.2ms
Speed: 10.8ms preprocess, 250.2ms inference, 1.7ms postprocess per image at shape (1, 3, 640, 640)
Total objects detected: 3
  person: 3
Saved annotated image as output.jpg

The real finding, not just an accuracy gap: COCO's 80 classes have almost nothing industrial — no forklift, pallet, crate, PPE-specific class, or factory equipment exists in the vocabulary at all. A pretrained general-purpose model is fundamentally blind to industry-specific objects, not just "less accurate" on them — the concept doesn't exist in its vocabulary. It's what directly motivated Levels 5 and 6.

Presentation framing: "Here are the 80 classes this model knows. Notice what's missing. That's the real gap between a general-purpose pretrained model and a production-ready solution for a specific factory floor."