AI & Machine Learning

Data Annotation (Data Labelling)

Data annotation, or labelling, is the process of adding correct answers to raw data, such as class labels, bounding boxes, segmentation masks or transcriptions, so that the data can be used to train and evaluate supervised machine learning models.

Annotation types depend on the task: image-level labels for classification, bounding boxes for object detection, pixel masks for segmentation, keypoints for pose estimation, text spans for entity extraction and transcripts for speech. Annotators follow written guidelines using labelling tools, and quality is controlled through review, consensus between annotators and agreement metrics. Model-assisted labelling and active learning reduce effort by pre-labelling data and prioritising uncertain examples.

In manufacturing, annotation is often the largest effort in a visual inspection project, because defects must be labelled consistently by people who understand the quality criteria. Defect catalogues and boundary samples agreed with quality teams help keep labels consistent.

Inconsistent or wrong labels limit achievable accuracy and distort evaluation. Guidelines should define edge cases, and test sets should be labelled with extra care. Confidential images and personal data in datasets require access control, and outsourcing labelling raises data protection considerations.

Key points

• Adds ground-truth labels for supervised learning
• Includes labels, bounding boxes, masks, keypoints and transcripts
• Consistency is controlled through guidelines, review and agreement metrics
• Often the largest effort in industrial vision projects

Related terms

• Supervised LearningAI & Machine Learning
• Object DetectionAI & Machine Learning
• Image SegmentationAI & Machine Learning
• Synthetic DataAI & Machine Learning
• AI BiasAI & Machine Learning

Terms that refer to Data Annotation (Data Labelling)

• Overfitting and UnderfittingAI & Machine Learning