Data labeling is the process of adding structured annotations to raw data so that machine learning models can learn from it. In computer vision, this means drawing bounding boxes around objects, tracing polygon outlines for segmentation masks, placing keypoints on joints for pose estimation, or assigning class tags for image classification. The quality of labels directly determines model quality — noisy, inconsistent, or missing annotations are one of the top reasons models underperform in production.

Labeling workflows range from fully manual (human annotators using tools like Datature Nexus, CVAT, or Label Studio) to model-assisted (a pre-trained model generates initial annotations that humans correct). Active learning takes this further by selecting which images to label next based on where the model is most uncertain, reducing the total labeling effort by 40-60% in typical projects.

At scale, labeling becomes a project management challenge. Teams need clear annotation guidelines, quality control checks (inter-annotator agreement, review queues), and version tracking. The cost of labeling varies widely: image classification tags take seconds per image, bounding boxes take 10-60 seconds each, and pixel-level segmentation masks can take several minutes per image.