Object detection answers two questions at once: what objects are in an image, and where are they located? The model predicts bounding boxes, class labels, and confidence scores for objects such as cars, people, invoices, tools, defects, or safety equipment.
Why it matters
Detection is used in retail analytics, quality inspection, autonomous driving, medical imaging, warehouse safety, sports analysis, document processing, and accessibility tools. It turns raw pixels into actionable locations.
Key terms
- Bounding box: rectangle around a detected object, usually represented by coordinates.
- Class label: predicted category such as person, vehicle, product, or defect.
- Confidence score: model estimate that a predicted box contains the class.
- IoU: Intersection over Union, a measure of how much two boxes overlap.
- NMS: Non-Maximum Suppression, a filtering step that removes duplicate boxes.
- Anchor boxes: predefined box shapes used by some detectors to predict objects.
- mAP: mean Average Precision, a common detection quality metric.
How a detector works
- A backbone network extracts image features.
- A detection head predicts candidate boxes, classes, and confidence values.
- Low-confidence boxes are filtered out using a confidence threshold.
- Overlapping duplicate boxes are removed with NMS using an IoU threshold.
- The remaining boxes become the final detections shown to users or downstream systems.
Visual explanation suggestion
Show an image-like grid with candidate boxes. A confidence slider fades weak boxes out, and an IoU/NMS slider removes overlapping duplicates so learners see why threshold choices change final detections.
Common mistakes
- Using only accuracy. Detection needs localization metrics such as IoU and mAP.
- Ignoring data labeling quality. Bad boxes teach the model bad boundaries.
- Choosing a confidence threshold without considering false positives and false negatives.
- Testing only clean images and missing lighting, blur, occlusion, camera angle, and small-object cases.
- Deploying a heavy model without measuring real latency on target hardware.
Interview-style questions
- What is the difference between classification, detection, and segmentation?
- How does Non-Maximum Suppression work?
- Why can mAP be more useful than accuracy for object detection?
- How would you debug a detector that misses small objects?
Related lessons
- Images as Data
- CNN - Convolution
- CNN - Pooling & Full Architecture
- Image Augmentation
- Edge Deployment & Optimization
Related project/template CTA
Practice this with the Object Detection Project template, then extend it with deployment and monitoring from the MLOps Starter Kit.