What does “Annotation” mean?
Human-added labels or metadata used to train or evaluate machine learning models.
How Annotation Works
Raw data: photos, text, audio
Human adds labels
Quality check by a second person
Labeled dataset
Used to train or test a model.
Practical Example
To train a model that detects stop signs, annotators draw boxes around each stop sign in thousands of street photos. A second reviewer checks a sample of boxes, and the labeled set is then used to train and test the detector.
Why It Matters
Many models are only as good as their labels. Careful, consistent annotation often matters more than adding more data.
Common Questions About Annotation
Q: Is annotation the same as data labeling?
They are closely related terms. Annotation often implies richer markup such as boxes or notes, while labeling often means assigning a category.
Q: Who does annotation?
Trained staff, contractors, domain experts such as doctors for medical data, or sometimes crowds of online workers.
Q: What makes annotation hard?
Ambiguous cases, unclear instructions, and fatigue can cause annotators to disagree, so teams write guidelines and measure agreement between people.