Learn AI /Computer Vision and Language Technology /Object detection
Object detection
- Intermediate
- Computer Vision and Language Technology
- Programme & data teams
- Technical readers
Back to the milestone you came from
In five seconds
What it is
A computer vision task that finds instances of chosen object types in an image or video and marks where each one appears.
In plain words
Object detection finds chosen kinds of things in a picture and draws a box around each one it finds.
An everyday example
Imagine receiving a photograph of a crowded kitchen counter and being asked to mark every apple and orange. You draw a box around each fruit, label it and note how confident you are. If you accidentally draw two overlapping boxes around the same apple, you keep the better one. An orange partly hidden behind a bowl may still be missed. Object detection performs this kind of locating and labelling automatically.
This analogy helps with finding and marking locations, but it can mislead in two ways. First, duplicate registry records are not visual objects in the same sense as people or vehicles, so the real task in computer vision depends on visual patterns in pixels rather than case content. Second, people on a table can use context and judgement in flexible ways, while a detection model only matches patterns it has learned from training data and may fail when lighting, angle, resolution or object type changes.
Remember thisObject detection identifies individual instances of selected object types and marks their locations, usually with bounding boxes and confidence scores.
Understand it
Understand it
Why it matters
For non-specialists, object detection matters because it changes what a system can tell you. A model that only classifies an image might say "there are vehicles here". A detector can say how many and where they are. Its outputs can also support systems that use tracking or other rules to estimate whether objects cross a line, enter an area or are missing from a scene. That makes it useful for counting, safety checks, inventory support, traffic analysis, document layout analysis and quality inspection. It also changes the risks: accuracy can vary by lighting, camera angle, object size, clutter and the training data used, so procurement, oversight and human review decisions should match the stakes.
How object detection turns an image into labelled boxes
After viewing this diagram, a reader should be able to explain that object detection takes an image, searches for candidate objects, removes duplicate overlaps, and returns labels plus locations.
01
1. Input image
The process starts with a picture or one frame from a video.
Draw a simple street-market scene with people, bicycles and boxes inside a photo frame.
02
2. Candidate detections
The model proposes many possible object locations and labels based on visual patterns.
Overlay several semi-transparent rectangles of different sizes around likely objects, including some incorrect and overlapping ones.
03
3. Filter overlaps
Very similar overlapping proposals are compared, and weaker duplicates are removed.
Fade out lower-confidence overlapping rectangles; keep one darker rectangle per object. Show small confidence tags such as 0.91 and 0.54.
04
4. Final output
The result is a shorter list of detected objects with labels, positions and confidence scores.
Show 3 to 5 final boxes with labels like person, bicycle and carton, plus a side list with label, box coordinates and confidence.
A left-to-right flow with four panels. Panel 1 shows an input photo. Panel 2 shows many faint candidate rectangles over likely objects. Panel 3 shows overlapping boxes being filtered so only the strongest remain. Panel 4 shows the final labelled boxes and a small results list. Read from left to right to follow the procedure from input to output.
Hovering or tapping a box highlights the matching row in the output list; a toggle shows all candidate boxes versus final boxes.
Describe this without the diagram
The diagram shows object detection as a four-stage flow. First, there is an input image containing several visible items. Second, the system places many possible rectangles over parts of the image where objects might be. Third, overlapping rectangles that refer to the same item are compared, and weaker duplicates are removed. Fourth, the remaining rectangles are shown with labels and confidence scores, along with a short table of results. The key idea is that the system does not only name object types; it also marks where each detected object appears.
With reduced motion: A static four-panel figure replaces animation. Panel 1 shows the original image. Panel 2 shows all candidate boxes at once. Panel 3 shows weaker overlapping boxes faded or crossed out. Panel 4 shows only the final labelled boxes and the results table.
Object detection takes an image or video frame as input. The system looks for specific categories such as people, vehicles, buildings or cartons. The output is usually a list of findings: the object label, a location, and a confidence score, which is a number indicating the strength the model assigns to that finding.
The location is often shown with a bounding box, a rectangle drawn around the detected object. This is what makes object detection different from image classification. Image classification answers a whole-image question such as "What is in this picture?" Object detection answers "What objects are here, and where are they?"
A common approach analyses many visual patterns, proposes possible object locations and labels, then removes weaker duplicate proposals. In video, detection can be repeated frame by frame, sometimes with tracking added afterwards to follow the same object over time.
Object detection is also different from image segmentation. Segmentation assigns labels at pixel level to outline objects or meaningful regions, while object detection usually gives a rectangle around each detected object. That rectangle is often good enough for counting, monitoring, search, safety alerts or checking whether something important is present in a scene.
How “Marking fruit in a photograph” maps onto the real thing
| In the story | In the AI system |
|---|---|
| each pile of files on the table | a possible object instance in the image |
| the coloured frame card around a pile | a bounding box showing location |
| the sheet with position and certainty | the output list of labels, coordinates and confidence scores |
| removing overlapping extra frame cards | suppressing duplicate detections |
A worked example
An annex to the handover note demonstrates the detector on one cleared test photograph of a furnished training room.
Starting from: The reference annotation contains 24 chairs and 6 tables, including one chair partly hidden behind a table.
- The detector produces candidate boxes across the image, including several overlapping boxes around some chairs.
- It assigns each candidate a class label and confidence score, such as chair 0.92 or table 0.87, and refines the box coordinates.
- Post-processing removes near-duplicate overlaps, retaining the strongest box when several correspond to the same object.
- The remaining detections are compared with the reference annotation. The model finds all 6 tables and 23 of the 24 chairs, missing the partly hidden chair.
Result: The output is a set of labelled, localised boxes with confidence scores. It supports counting and review, but the comparison shows that an apparently precise total can still be incomplete when objects overlap.
In a United Nations workplace
A statistician in an evaluation office is finalising a handover note before close of business. It records how an approved detector was tested on photographs attached to implementing partner monitoring reports, locating chairs and tables in training rooms so the incoming colleague can reproduce the review rather than inherit a folder called "final-final".
The detector returns a label, location and confidence score for each visible item. This can accelerate an initial review of a large photo set, while sampled manual checks show whether crowding, unusual camera angles or locally common furniture are affecting the results.
Before you try thisMonitoring photographs may contain faces, name badges, location details or other sensitive information. They must not be uploaded to a tool that has not been approved for that material. Use cleared images, minimise or redact personal data, and verify detections against the underlying records. A visible chair is not proof that it was procured, delivered as reported or usable.
Where you may hear it
- During a vendor demo to an ICT governance board
- In a procurement unit reviewing requirements for image-analysis software
- At an evaluation office meeting on analysing monitoring photographs
- In an inter-agency working group discussing satellite or drone imagery
Avoid confusion
Avoid confusion
Often confused with
Image classification
A system assigns one or more predefined labels to an image based on patterns in its pixels.
Image segmentation
Image segmentation divides an image into pixel-level regions; semantic segmentation labels them as roads, buildings or water.
Facial recognition
A computer vision method that compares face images or representations to estimate whether they show the same person.
Bounding box
A rectangular marker that shows where an object or text region appears in an image, usually by its position, width and height.
Common misconceptions
Common misconception: If a box appears around an object, the system has identified it precisely.
Correction: A bounding box is only an approximate location. It often includes background, may miss part of the object, and can shift when the object is small, partly hidden or viewed at an unusual angle. If exact shape matters, image segmentation may be more suitable.
Common misconception: Object detection simply counts things, so it should always get the total right.
Correction: Counting is a downstream use of detection, not the same task. Missed detections, duplicate boxes, poor visibility and crowding can all distort counts. A detector can support counting, but the count is only as good as the detection quality and the review process.
Common misconception: A high confidence score means the result is probably true in any setting.
Correction: Confidence scores are model outputs, not guarantees. A score can look high even when the image comes from a different camera, country, season or environment than the training data. Scores need calibration and validation in the real deployment context.
Quick check
Which output best matches object detection rather than image classification?
Object detection returns object instances plus locations, usually as bounding boxes with labels and confidence scores. A single label is image classification, a paragraph is image captioning, and pixel-level masks are closer to image segmentation.
Go deeper
Go deeper
The simple version
Object detection finds chosen kinds of things in a picture and draws a box around each one it finds.
The technical version
Object detection is a computer vision task that predicts zero or more object instances in an image, typically as class labels plus localised coordinates such as bounding boxes and confidence scores. In a common supervised setup, detectors learn visual features from annotated training images. One common pipeline generates candidate regions or dense predictions, classifies them, refines box coordinates, and applies post-processing to remove near-duplicate overlaps. Evaluation usually considers both localisation and classification, often through overlap-based criteria rather than label accuracy alone. The term differs from image classification, which predicts labels for the whole image, and from image segmentation, which assigns labels at pixel level instead of only rectangular extents.
Reviewed external resources
EncyclopaediaObject detection (opens in a new tab)Wikipedia
This link gives a broad overview of object detection concepts and history, which suits readers who want a general starting point.
Intermediatelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This glossary link suits readers who want short definitions of related machine learning terms without reading a long article.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course link is useful for learners who want a structured introduction to machine learning ideas that support detection systems.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Object detection (opens in a new tab)YouTube search
This search results page helps readers find video explanations of object detection in different teaching styles, which suits visual learners; it is not a single curated video.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Object detection (opens in a new tab)Reddit search
This search results page helps readers browse community questions and practical discussions about object detection, which suits readers comparing real-world experiences; it is not a single authoritative source.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Object detection identifies individual instances of selected object types and marks their locations, usually with bounding boxes and confidence scores.