Annotation services

Bounding box annotation services

Bounding box annotation is drawing the smallest axis-aligned rectangle that contains each object of interest in an image and tagging it with a class, such as car, person or defect. The boxes and classes become training data for object detection models like YOLO, Faster R-CNN and DETR.

Trains
Object detection
From
$0.057 / label
Review
Every job
Formats
YOLO TXT

Is it the right label?

When to use bounding boxes

Object detection — the model learns to find and classify every object in a new image.

  • You need to know what is in the image and roughly where — counting, presence checks, tracking-by-detection.

  • Objects are compact and roughly rectangular from the camera's view: vehicles, people, packages, products.

  • You need a large dataset fast. Boxes are the quickest spatial label to produce and review.

When something else fits better

If you need the object's exact shape or area — a crack, a tumour, a crop row — or the objects are long, thin or diagonal, a box carries too much background. Use polygon segmentation or oriented boxes instead.

Oriented boxes

How we label it

Our process

  1. 01

    Agree the box rule

    Tight to visible pixels or to the estimated full extent? Minimum object size? How to treat reflections and groups? Written into the guideline before labeling starts.

  2. 02

    Label every instance

    Annotators box every instance of every class in the frame — not just the obvious ones — with occlusion and truncation flags where your guideline needs them.

  3. 03

    Model-assisted where it helps

    On large runs a detector pre-labels and annotators correct it. Every box is still checked by a person; pre-labels never ship unreviewed.

  4. 04

    Review and rework

    A reviewer checks each job against the guideline. Missed objects and loose boxes go back to the annotator, not into your dataset.

Quality

What our QA checks

Every job is reviewed by a person other than the annotator, on top of automated checks. Work below the agreed threshold goes back for rework.

Missed-object audits — the most expensive detection error is the box that was never drawn

Box tightness measured as IoU against reviewer-corrected boxes

Class confusion review on look-alike classes (van vs truck, helmet vs cap)

Structural checks: zero-area boxes, boxes outside the image, duplicate boxes

Pricing & delivery

Rates and formats

AnnotationRate
Object detection$0.057 / label

QA review is included. Bulk discounts apply to large projects; the final quote is confirmed after we review a sample of your data. Full pricing

Delivered as

YOLO TXTCOCO JSONPascal VOC XMLCVAT XMLDatumaro

Used for

Object detection · Retail shelf audits · Vehicle and traffic counting · PPE and safety monitoring · Package and pallet detection

Bounding boxes FAQ

Bounding boxes, answered

What is bounding box annotation?

Bounding box annotation is drawing a tight rectangle around each object in an image and labeling it with a class. It is the standard training data for object detection models such as YOLO.

How tight should a bounding box be?

The box should touch the outermost visible pixels of the object on all four sides, with no gap and no clipped edges. Whether to box only the visible part or the estimated full extent of an occluded object is a guideline decision we agree with you before labeling.

How are occluded and truncated objects labeled?

By default the visible part is boxed and the object is flagged as occluded or truncated, so you can filter or weight those labels during training. If your model needs amodal boxes (the full estimated extent), we label that way instead.

Which format do you deliver bounding boxes in?

YOLO TXT (normalised center-x, center-y, width, height), COCO JSON (x, y, width, height in pixels), Pascal VOC XML (xmin, ymin, xmax, ymax), CVAT XML or Datumaro — or a custom schema.

How much does bounding box annotation cost?

Bounding boxes are priced per box. The total depends mostly on how many objects each image contains, so we confirm the price after reviewing a sample of your images.

Need bounding boxes on your data?

Send us a few sample images or clips. We'll agree the guideline with you and send a plan and a per-label quote.