Image annotation glossary
Image annotation borrows its vocabulary from computer vision, data operations, and accessibility, which is why the same idea often has three names. This glossary gives one working definition for each term, plus how it relates to the automatic descriptions this tool produces.
Terms are listed alphabetically. Each entry is short on purpose: enough to answer a search or an AI query without a detour.
By Automatic Image Annotation editorial team · Updated Nov 1, 2026
active learning
active learning loop
Active learning selects the examples a model is least confident about and sends those for human review first. Annotation effort goes where it changes the model most, instead of being spread evenly across a dataset.
In this product — The same idea drives batch review here: you check generated descriptions before they reach a page, not every pixel behind them.
AI annotation
machine annotation · model-assisted annotation
AI annotation is annotation performed or assisted by a model rather than by a person alone. It covers vision-language models that write descriptions and detectors that place bounding boxes, and is usually paired with human review.
In this product — Every description on this site is AI annotation: a model writes the text, and you decide whether to publish it.
alt text
alternative text · alt attribute
Alt text is the short description stored in an image's alt attribute. Screen readers announce it, browsers show it when an image fails to load, and it is the minimum a page needs to stay accessible.
In this product — Alt text is one of the three outputs this tool returns for each image, alongside a caption and a longer description.
annotation pipeline
labeling pipeline
An annotation pipeline is the end-to-end flow from raw data to reviewed labels: ingestion, pre-labeling, human review, quality checks, and export. Pipelines differ mainly in how much human review they keep.
In this product — Here the pipeline is deliberately short — upload, generate, review, copy or export — because the output is page copy rather than training data.
annotation tool
labeling tool · image annotation software
An annotation tool is the interface used to create or review labels. Dataset tools provide box, polygon, and mask editors; content tools provide caption fields. The two are rarely the same product.
In this product — This site is the content-team kind of annotation tool: you review text, not draw boxes.
auto annotation
auto-annotation · automatic annotation
Auto annotation is shorthand for automatic annotation, most often used for pre-labeling: a model proposes labels that a person corrects. It cuts the cost of building a dataset without removing the reviewer.
In this product — Auto annotation here means the model drafts alt text, a caption, and a description; the review step is yours.
automatic image annotation
automatic image labeling · automated image annotation
Automatic image annotation is the process of generating descriptions, labels, or captions for an image with a model instead of a person. Depending on the dataset, the output is free text, a fixed label set, or coordinates such as bounding boxes.
In this product — This is what the product does: one uploaded image becomes alt text, a caption, and a wall label you can copy or export.
automatic labeling
automated labeling
Automatic labeling generates labels without manual input. In practice it usually means model-assisted labeling, where predictions are accepted, corrected, or rejected by a reviewer before they are trusted.
In this product — Descriptions are generated automatically, then reviewed in the workspace before you publish them.
bounding box
bbox · bounding rectangle
A bounding box is a rectangle drawn around an object, stored as x, y, width, height or as two corner points. It is the standard output format for object detection datasets such as COCO and YOLO.
In this product — Boxes are dataset output, not page copy. This tool writes text descriptions, so it intentionally does not draw boxes.
data annotation
dataset annotation
Data annotation is the broader discipline of labeling training data across images, text, audio, and video. In computer vision it covers classification, object detection, and segmentation.
In this product — The tool produces publishable text rather than training labels, but the underlying image understanding overlaps.
data labeling
dataset labeling
Data labeling is often used interchangeably with data annotation, especially in machine learning operations, where it names the step that produces ground truth for supervised models.
In this product — If you need labeled training data, you want a dataset platform; if you need descriptions for pages, that is what this tool produces.
ground truth
reference labels
Ground truth is the human-verified reference data that a model's predictions are measured against. Its quality sets the ceiling for accuracy metrics, because a model can only be as correct as the answer key it is graded on.
In this product — For published copy, the reviewer is the ground truth: your correction is the final answer, not the model's draft.
human-in-the-loop
HITL · human review
Human-in-the-loop keeps a person in the workflow to review, correct, or approve machine output before it is published or used as training data. It is the usual compromise when full automation is too risky.
In this product — Every result stays editable in the workspace and in your history, so a person always approves the final text.
image annotation
picture annotation · image tagging
Image annotation is the act of attaching information to an image: a category label, a caption, a bounding box, or a segmentation mask. It is the umbrella term for every task that turns pixels into text or structured labels.
In this product — This site covers the text branch of image annotation: alt text, captions, and descriptions.
image caption
photo caption · caption text
An image caption is a sentence that describes an image, used under photos, in CMS fields, and in social posts. It is usually longer and more contextual than alt text and can carry tone.
In this product — Captions are the second output of each run and are meant for visible page copy rather than the alt attribute.
image description
image alt description · long description
An image description is a longer, structured account of what an image shows, including subjects, actions, visible text, and context. It is the most complete text output of an annotation system.
In this product — The description is the longest of the three outputs and is the best fit for CMS body fields.
image labeling
image labelling · picture labeling
Image labeling assigns one or more predefined labels to an image, usually for classification. Unlike captioning, the output is a fixed vocabulary rather than free text, which makes labels easy to count, filter, and compare.
In this product — Labels and free text answer different questions. This tool returns free text because pages need sentences, not categories.
polygon annotation
polygonal annotation · polygon mask
Polygon annotation traces an object's outline with a closed set of points. It is more precise than a bounding box for irregular shapes and is common in segmentation datasets.
In this product — Not applicable to text output: a paragraph has no outline to trace.
semantic segmentation
pixel segmentation · segmentation mask
Semantic segmentation assigns a class to every pixel, producing a mask per category rather than a box per object. Instance segmentation goes further and separates individual objects of the same class.
In this product — Segmentation is a dataset task. The practical link is that good segmentation models often power the image understanding behind description tools.
weak supervision
weakly supervised learning
Weak supervision builds training labels from noisy, indirect sources such as heuristics, existing metadata, or model predictions, instead of hand-labeling every example. It trades label precision for volume.
In this product — Descriptions are generated from the image itself with no external metadata, so the text does not depend on weak or inherited labels.
Keep reading
Two pages that go deeper than a definition.
Turn the definition into a result
Upload an image and get alt text, a caption, and a description in one run.