Object Detection — Level 5: YOLO-World

Approach

Ultralytics YOLOWorld (yolov8s-world.pt) — the direct answer to Level 4's vocabulary gap.

Detects any object described in plain text at inference time — no training or fine-tuning required, unlike COCO's fixed 80-class list.

model = YOLOWorld("yolov8s-world.pt")
model.set_classes(["forklift", "cardboard box", "pallet", "helmet", "person"])
results = model.predict(image_path, conf=0.08, iou=0.5)

Getting there

Worth showing as process, not just the end result.

Initial results only found "person" — the default confidence threshold (0.25) was filtering out less-common/compound terms scoring lower; lowering to ~0.08 revealed the rest.
"box" matched any box-shaped object — electrical cabinets, mystery equipment — too generic a term. This matches language patterns, not fixed categories.
Refined to specific phrases: "cardboard box," "wooden crate," "electrical cabinet," "pallet jack," "safety vest," "helmet."
Raising imgsz to 1280 (from the default 640) recovered detection of smaller/distant objects (a pallet jack handle) lost to downscaling, at the cost of needing to re-lower conf since scores shifted slightly.

Original, unannotated warehouse photo with two workers in safety vests and helmets — **Before** — the original, unannotated photo.

YOLO-World correctly detecting person, helmet, electrical cabinet, and pallet jack in the same warehouse photo — **After** — person, helmet, electrical cabinet, and pallet jack all correctly detected with distinct per-class colors and confidence scores.

Terminal output

image 1/1 warehouse1.jpg: 864x1280 2 helmets, 4 persons, 2 electrical cabinets, 1 pallet jack, 1251.3ms
Speed: 20.3ms preprocess, 1251.3ms inference, 1.5ms postprocess per image at shape (1, 3, 864, 1280)
warehouse1.jpg -> Total objects detected: 9
  person: 4
  helmet: 2
  electrical cabinet: 2
  pallet jack: 1
Saved annotated image as output1.jpg

A synthetic, 3D rendered forklift warehouse image with zero detections drawn — Not every scene worked: this synthetic, 3D-rendered forklift warehouse image returned zero detections (shown as originally captured — with nothing detected, there's no annotated version to compare against) — see Level 6 for what changed on the exact same image.

Terminal output

image 1/1 warehouse3.jpg: 736x1280 (no detections), 1527.6ms
Speed: 23.0ms preprocess, 1527.6ms inference, 1.5ms postprocess per image at shape (1, 3, 736, 1280)
warehouse3.jpg -> Total objects detected: 0
  No objects detected - try lowering conf further or rephrasing classes.
Saved annotated image as output3.jpg

Open-vocabulary detection matches language patterns, not fixed definitions. Word choice directly and significantly changes what gets detected — a real UX/deployment consideration, not just a modeling detail.