Object detection is a computer vision task that locates and classifies objects in an image or video, drawing a bounding box around each one and assigning it a class. It answers both what is present and where, unlike classification, which only says what.
Object detection combines two questions into one task: what objects are in the image, and where is each one. A detector outputs a set of bounding boxes, each with a class label and a confidence score, so a single image yields many detections.
It sits between image classification, which labels the whole image with no location, and segmentation, which labels exact pixels: detection gives coarse location, a box, for every object, which is enough for a huge range of applications, counting, tracking, and triggering actions. Its training data is bounding-box annotation, the boxes a model learns to predict.
Key takeaways
Object detection finds and classifies each object, outputting a box, a class, and a confidence.
It is the middle ground, more than classification, which has no location, less than segmentation, which has exact shape.
It learns from bounding-box annotation and is scored with IoU and mean average precision (mAP).
What object detection provides
How detection compares to classification and segmentation.
How detection compares to classification and segmentation.
Task
Output
Location
Image classification
A class for the image
None
Object detection
Boxes with classes
Coarse, boxes
Segmentation
Per-pixel classes
Exact
How it works
A detector proposes boxes and classifies them, keeping those above a confidence threshold. In FiftyOne, detections render as boxes with labels and scores, and evaluation matches predictions to ground truth by IoU and computes mAP, so you can find false positives, missed objects, and the confidence threshold that works best.
Why it matters
Object detection is one of the most widely deployed CV tasks, the workhorse behind surveillance, retail analytics, driving, and inspection. Detection metrics hide where models fail. mAP is an average over classes and thresholds, so a model can post a strong number while quietly missing small objects, rare classes, or crowded scenes. The useful move is to slice evaluation, by object size, class, and scene density, rather than trusting the headline mAP, because that is where you find the failures that matter and the data you need to annotate next.
Frequently asked questions
What is the difference between object detection and image classification?
Classification labels the whole image. Detection also locates each object with a box.
What is the difference between detection and segmentation?
Detection gives a box per object. Segmentation gives exact per-pixel shape.
How is object detection evaluated?
With IoU to match boxes and mean average precision (mAP) across classes and thresholds.