10.3.1 Object Detection Roadmap: Class plus Box

Object detection adds location to classification: what object is present, and where is it in the image?

See the Box Workflow First

The important concepts are bounding box, class, confidence, IoU, threshold, false positive, false negative, and mAP.

Run an IoU Check

IoU measures how much the predicted box overlaps the ground-truth box.

truth = (10, 10, 50, 50)
pred = (20, 20, 60, 60)

def area(box):
    x1, y1, x2, y2 = box
    return max(0, x2 - x1) * max(0, y2 - y1)

ix1 = max(truth[0], pred[0])
iy1 = max(truth[1], pred[1])
ix2 = min(truth[2], pred[2])
iy2 = min(truth[3], pred[3])
intersection = area((ix1, iy1, ix2, iy2))
union = area(truth) + area(pred) - intersection

print("iou:", round(intersection / union, 3))

Expected output:

iou: 0.391

Detection debugging starts by printing boxes and metrics. Do not judge detection quality from one nice screenshot.

Learn in This Order

Step	Read	Practice Output
1	Detection overview	Explain box, class, confidence, IoU, mAP
2	Classic detectors	Compare two-stage and one-stage ideas
3	YOLO	Understand grid prediction, threshold, NMS, and speed trade-offs
4	Detection practice	Record false positives, missed detections, and threshold changes

Evidence to Keep

Keep this page’s proof of learning as a small evidence card:

Input Image

detection sample with ground-truth or expected objects

Prediction

boxes, labels, confidence scores, IoU, and threshold settings

Metric

precision/recall, mAP, false positives, and false negatives

Failure Check

small object, overlap, NMS, poor labels, or confidence threshold

Expected Output

annotated image plus detection metrics or error buckets

Pass Check

You pass this chapter when you can explain a detection result with boxes, confidence, IoU, and at least one false-positive or false-negative case.

Check reasoning and explanation

1. A passing answer maps the task to the right visual output: class label, bounding box, mask, OCR text, embedding, or video event.
2. The evidence should include a rendered visual artifact and one metric or qualitative error note.
3. A good self-check names one visual failure mode such as class confusion, missed objects, bad masks, lighting shift, domain shift, or weak annotation quality.