Object Detection
Also known as: visual object detection, image object localization, region-based object detection, bounding-box detection
Object detection is a computer vision task in which a deep neural network simultaneously locates and classifies every instance of one or more object categories within an image, producing a bounding box and a class label for each detected object. Modern detectors — from the R-CNN family to YOLO and DETR — achieve near-human accuracy at real-time speeds on standard benchmarks.
Key highlights
- State-of-the-art accuracy on benchmark datasets, often matching or exceeding human performance on specific categories.
- Simultaneous localisation and classification in a single forward pass with modern one-stage detectors.
- Rich ecosystem of pre-trained models (COCO, Open Images) enabling effective transfer learning with small domain datasets.
- Scalable to video streams and real-time applications with optimised architectures such as YOLOv8.
- Flexible: handles varying numbers of objects per image and objects at multiple scales through feature pyramid networks.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use object detection when your research or application requires locating and classifying multiple object instances within images — for example, counting cells in microscopy, detecting pedestrians in video, or identifying products on shelves. It requires annotated training images with bounding boxes, which is laborious to produce; if annotation budget is very limited (fewer than a few hundred images), strongly consider transfer learning from a pre-trained detector. Do not use object detection if you only need image-level labels without spatial localisation (use image classification instead), or if you need pixel-level boundaries (use instance segmentation instead).
Strengths & limitations
- State-of-the-art accuracy on benchmark datasets, often matching or exceeding human performance on specific categories.
- Simultaneous localisation and classification in a single forward pass with modern one-stage detectors.
- Rich ecosystem of pre-trained models (COCO, Open Images) enabling effective transfer learning with small domain datasets.
- Scalable to video streams and real-time applications with optimised architectures such as YOLOv8.
- Flexible: handles varying numbers of objects per image and objects at multiple scales through feature pyramid networks.
- Requires bounding-box annotated training data, which is expensive and time-consuming to produce at scale.
- Computationally intensive — training large detectors demands multi-GPU hardware and significant memory.
- Performance degrades on heavily occluded objects, very small objects, and densely packed scenes.
- Hyperparameter sensitivity: anchor sizes, confidence thresholds, and NMS parameters all require domain-specific tuning.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
Which detector should I choose — Faster R-CNN, YOLO, or DETR?
Faster R-CNN generally achieves higher accuracy but is slower; it is a good choice when precision matters more than speed. YOLO variants (v5, v8) are optimised for real-time inference and are easier to deploy on edge hardware. DETR and its successors offer competitive accuracy without NMS but need more training data and longer training times.
How many annotated images do I need to train a detector?
From scratch you typically need thousands of annotated images per class. With transfer learning from COCO-pretrained weights, useful results can sometimes be achieved with a few hundred images per class, but performance is highly dataset- and task-dependent; always validate on a held-out test set.
What is mAP and why is it the standard metric?
Mean Average Precision (mAP) summarises the precision-recall curve for each class at a given IoU threshold and averages it over all classes. Unlike accuracy, it accounts for both localisation quality (via IoU) and classification confidence, making it far more informative for the class-imbalanced, multi-object detection task.
Can object detection work on video?
Yes. Frame-by-frame detection with a fast model like YOLO is common, often combined with a tracker (e.g., SORT, DeepSORT) to assign consistent IDs to objects across frames and reduce flickering detections.
When should I use instance segmentation instead?
When you need pixel-accurate object masks — for example, to measure exact object boundaries, compute areas, or handle highly irregular shapes — use instance segmentation (e.g., Mask R-CNN). Object detection with bounding boxes is sufficient when approximate location and count are the research goal.
Sources
- 1.Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014). Rich feature hierarchies for accurate object detection and semantic segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 580–587.
- 2.Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You Only Look Once: Unified, Real-Time Object Detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 779–788.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Object Detection. ScholarGate. https://scholargate.app/deep-learning/object-detection