What is an annotation format?
An annotation format is the convention for encoding image labels in files that training tools can read. Each label records its image, the object’s class, and its geometry: a bounding box, a polygon, a pixel mask, or keypoints. The format decides how those values are written and how they are linked to the image, usually by file name.
Three families are common. COCO (Common Objects in Context) stores all labels for a dataset in one JSON (JavaScript Object Notation) file. Pascal VOC (Visual Object Classes) writes one XML (Extensible Markup Language) file per image. The YOLO (You Only Look Once) format uses one text file per image, with a class number and normalized box coordinates on each line. Small projects often start with a CSV (comma-separated values) file and convert later.
Why annotation formats matter
The format is the handoff point between in-house labelers, outsourced labeling vendors, and the training framework. A collections agency that labels scanned remittance forms in one tool and trains in another needs a conversion step, where coordinates get mirrored, class IDs shift, or unlabeled images get dropped. Annotation files deserve the same version control, data lineage, and quality checks as any other dataset.
What to watch for
- Confirm the coordinate convention before converting, and spot-check converted boxes visually afterwards.
- Ask whether a vendor’s tool exports to an open format; labels locked in a proprietary format are a form of lock-in.
- Keep the class list in one controlled file so that class 3 means the same thing in every batch.