Understand Object Detection
Boxes, classes, and confidence — what detection actually returns and what it does not.
Object detection is the workhorse. Most production CV systems are detection systems with extra logic on top — counters, trackers, alerters. If you understand exactly what a detection model returns and what it doesn't — class, box, and confidence — you can build almost anything.
Hands-on
What comes out of the model#
A detection model returns a list. Each item has:
- Class — an integer index that maps to a name (e.g.
0 → person,2 → car). - Box — four numbers:
x1, y1, x2, y2(top-left and bottom-right pixel coordinates). - Confidence — a number between 0 and 1 measuring how sure the model is.
That is the whole API. Everything else — counting, alerting, tracking, masking sensitive regions — is code you write that consumes this list.
results = model("warehouse.jpg")
for box in results[0].boxes:
cls = int(box.cls[0]) # class index
conf = float(box.conf[0]) # confidence
x1, y1, x2, y2 = box.xyxy[0] # corners
print(f"{model.names[cls]} @ {conf:.2f}: ({x1:.0f},{y1:.0f}) → ({x2:.0f},{y2:.0f})")Confidence is not class probability#
A common confusion: confidence ≠ "the model is X% sure this is a forklift." It's closer to "the model is X% sure something is here and its best guess for the class is forklift." Most modern anchor-free detectors — including YOLO26 — bundle the two into one score; for some older models they're separate.
The practical consequence: a single confidence threshold may be too coarse if your classes differ in difficulty. You can apply different thresholds per class — see Predict mode for the full list of inference arguments.
Duplicate boxes and NMS#
Detectors propose many candidate boxes per object. Non-Maximum Suppression (NMS) removes overlapping duplicates of the same class, keeping the highest-confidence one. NMS has one important knob, the IoU threshold, that decides "how much overlap counts as a duplicate."
If your model misses one of two objects standing next to each other, the bug may not be the model — it may be NMS deduplicating two real objects. Lower the IoU threshold (e.g. iou=0.5 → iou=0.7) and the second one comes back.
Two ways to be wrong#
A detection is wrong in two distinct ways:
| Failure | What's broken | Fix |
|---|---|---|
| Right class, bad box | Localization | More data, larger imgsz, or rethink the labeling rules for tight crops |
| Wrong class | Classification | More examples of the confused class, or a class merge if the distinction isn't real |
A useful diagnostic loop is to look at low-confidence true positives and high-confidence false positives. They tell you very different things.
Picking a confidence threshold#
The default conf=0.25 is sane but rarely optimal. Pick it deliberately based on which mistake is more expensive:
- Safety / alerts: false negatives are expensive — accept some false positives — lower the threshold. This is exactly the tradeoff in warehouse safety monitoring.
- Search / recommendation: false positives are visible to users — raise the threshold.
- Counting / metrics: balance — sweep a range and pick the threshold that minimizes count error on your validation set.
Show solution
from ultralytics import YOLO
model = YOLO("yolo26n.pt")
results = model("https://ultralytics.com/images/bus.jpg", conf=0.25, iou=0.7)
for box in results[0].boxes:
print(model.names[int(box.cls)], float(box.conf), box.xyxy.tolist())