What is bounding box annotation?
A bounding box is a rectangle drawn around an object in an image or a video frame, with a label that says what the object is. It is the most widely used kind of annotation for object detection: a model learns from thousands of labeled boxes where objects are and what they are called.
Boxes are quick to draw and simple to check. That makes them the practical choice when you need the position and class of an object, but not its exact outline.
What we label
- 2D boxes around objects in photos and video frames.
- Rotated boxes for objects that sit at an angle, such as vehicles seen from above or text in a street scene.
- 3D cuboids that capture depth and direction, in images and in LiDAR point clouds.
- Attributes on every box: class, occlusion, truncation and any property your model needs, such as colour, state or direction of travel.
- Tracking of the same object across video frames with one consistent ID.
Where bounding boxes are used
- Driving and driver assistance: vehicles, pedestrians, signs and traffic lights.
- Retail and e-commerce: products on shelves and in catalog photos.
- Logistics: parcels, pallets and shipping labels.
- Safety: people, helmets and protective equipment on a site.
- Agriculture: fruit, plants and livestock, counted and located.
- Documents: tables, signatures, stamps and form fields.
How we keep boxes consistent
Drawing one box is easy. Drawing a hundred thousand the same way, across many annotators, is the real work. A project runs in four steps:
- Guidelines first. We agree the class list and the edge cases with you: how tight a box is, how to treat hidden or cut-off objects, and the smallest object worth labeling.
- Pilot batch. A small sample is labeled and reviewed with you, and the guidelines are corrected before volume work starts.
- Production with review. Work is checked in a second pass, and errors go back to the annotator with the reason.
- Delivery. You receive the labels in the format your training pipeline uses, such as COCO, YOLO or Pascal VOC.
Quality you can measure
Pripton's AWS Marketplace listings report 98.5% average accuracy, with 500+ trained annotators and more than 60,000 images annotated a day. If a project falls below 96.5% accuracy, we rework it.
Box, polygon or segmentation?
- Choose boxes when the model needs to find and classify objects.
- Choose polygons or image segmentation when the exact outline matters, for example to measure area.
- Choose keypoints when the model needs pose or the position of specific parts.
Not sure which fits your model? Our image annotation team can label a small sample each way so you can compare.
Start a project
Send a few sample images and your class list through the contact page. If you buy through AWS, our listings are on the platform page.
