Object Detection is a Computer Vision task that identifies instances of objects and estimates their locations within an image or video frame. Unlike image classification, which produces image-level labels, object detection can identify several objects and indicate where each one appears.
A common representation of a detection is a bounding box, usually described by coordinates and dimensions, together with a class label and a confidence score.
An object detection system typically performs these operations:
1. Image input: Receives an image or video frame.
2. Feature extraction: Learns visual patterns using a backbone network such as a CNN or Vision Transformer.
3. Object prediction: Estimates object categories and bounding-box coordinates.
4. Confidence filtering: Removes predictions with confidence scores below a selected threshold.
5. Duplicate removal: Uses methods such as Non-Maximum Suppression (NMS) to reduce overlapping duplicate detections.
6. Output: Returns object labels, locations, and associated confidence scores.
Object detection architectures are commonly grouped into two broad categories:
- Two-stage detectors: First generate candidate object regions and then classify and refine them. Faster R-CNN is an important example.
- One-stage detectors: Predict object locations and classes directly in a more unified process. YOLO and SSD are well-known examples.
Detection systems are evaluated using metrics such as Intersection over Union (IoU), precision, recall, and Average Precision (AP). IoU measures the overlap between a predicted bounding box and a reference bounding box. Mean Average Precision (mAP) summarizes average precision across classes under a specified evaluation protocol.
Object detection can be difficult when objects are small, partially hidden, densely packed, or visually similar to their surroundings. Changes in lighting, camera angle, and image quality can also reduce accuracy.
Object Detection
Object Detection is a Computer Vision task that identifies objects in an image or video and determines their locations, commonly using bounding boxes and class labels.
IoU = Area(B_pred ∩ B_true) / Area(B_pred ∪ B_true)
Explanation
Variables
| Symbol | Meaning |
|---|---|
B_pred | Predicted bounding box |
B_true | Ground-truth bounding box |
∩ | Intersection of two regions |
∪ | Union of two regions |
IoU | Intersection over Union overlap measure |
Example or application
A road-monitoring camera uses an object detection model to identify cars, buses, motorcycles, and pedestrians, returning a bounding box and class label for each detected object.
Text
Object detection combines object localization and classification. Detection models identify multiple objects, assign category labels, and estimate their positions within an image.
Key Points
Object detection identifies objects and their locations.
Bounding boxes commonly represent object locations.
Detectors produce class predictions and confidence scores.
One-stage and two-stage detectors use different prediction strategies.
Non-Maximum Suppression can remove duplicate overlapping detections.
IoU measures overlap between predicted and reference boxes.
Small, crowded, and partially hidden objects create challenges.
Applications
Autonomous vehicle perception
Traffic monitoring
Industrial inspection
Retail analytics
Robotics
Wildlife monitoring
Sports video analysis
Learning Outcomes
Define object detection.
Differentiate classification and localization.
Explain the role of bounding boxes and confidence scores.
Compare one-stage and two-stage detectors.
Understand IoU and basic detection evaluation.
Practice Questions
What is object detection?
How does object detection differ from image classification?
What information does a bounding box provide?
What is the purpose of Non-Maximum Suppression?
How is Intersection over Union calculated?
Related Concepts
Image Classification
Image Segmentation
Convolutional Neural Networks
Bounding Boxes
Intersection over Union
YOLO
Common Mistakes
Confusing image-level classification with object localization.
Treating confidence scores as guaranteed probabilities of correctness.
Assuming every detector uses the same architecture.
Ignoring the evaluation threshold when comparing IoU or mAP results.
Sources and verification
Sources are shown as provided in the project data. Not every source has yet been independently verified. Please report errors or missing citations.
- Ren, S. et al., Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks, NeurIPS, 2015.
- Redmon, J. et al., You Only Look Once: Unified, Real-Time Object Detection, CVPR, 2016.
Suggest a correction or source · More in Artificial Intelligence & Machine Learning