Data Annotation Glossary

Annotation terms, A to Z of the workflow

Data annotation

Adding labels to raw images, video, text, audio or 3D data so a machine learning model can learn from examples.

Bounding box

A rectangle drawn tightly around an object to mark its location and class in an image or video frame.

Image annotation Guide: bounding box vs polygon vs segmentation

Polygon annotation

An outline made of connected points that follows an object's true shape more closely than a box.

Semantic segmentation

Assigning a class to every pixel in an image, without separating individual objects of the same class.

Guide: semantic vs instance vs panoptic segmentation

Instance segmentation

Pixel-level masks that separate each individual object, so two touching cars get two different masks.

Guide: semantic vs instance vs panoptic segmentation

Panoptic segmentation

A combination of semantic and instance segmentation: every pixel gets a class, and countable objects also get an instance ID.

Guide: semantic vs instance vs panoptic segmentation

Keypoints

Named points placed on landmarks such as joints, facial features or hand knuckles, often linked into a skeleton.

Guide: keypoint annotation for pose estimation

Object tracking

Following the same object across video frames with one persistent track ID, including through occlusion.

Video annotation

Occlusion and truncation

Occlusion is when an object is partly hidden by another; truncation is when it is cut off by the frame edge. Written rules decide how each is labelled.

IoU (Intersection over Union)

The overlap between two shapes divided by their combined area, used to compare an annotation with a reference label.

tIoU (temporal IoU)

IoU applied to time: the overlap between two action segments divided by their combined duration, used to check action boundaries.

Inter-annotator agreement (IAA)

A measure of how consistently different annotators label the same data, used to spot unclear rules and guide adjudication.

Guide: inter-annotator agreement explained

Adjudication

Resolving disagreements between annotators by a reviewer decision that is recorded and fed back into the guidelines.

Egocentric video

First-person footage recorded from a head-mounted or wearable camera, showing the wearer's hands and the objects they use.

Egocentric video annotation

LiDAR

A sensor that measures distance with laser pulses and produces a 3D point cloud of the surrounding scene.

LiDAR & 3D annotation Guide: LiDAR point cloud annotation

3D cuboid

A 3D box placed in a point cloud with a position, dimensions and yaw (heading) to mark an object in space.

Sensor fusion

Combining calibrated camera and LiDAR data so a 3D label can be projected onto the matching 2D image and checked for consistency.

Named entity recognition (NER)

Labelling spans of text such as people, organisations, amounts or products with their entity type and exact character offsets.

Text annotation Guide: text annotation for NLP

Speaker diarization

Marking who spoke when in an audio recording, including overlapping speech and consistent speaker labels.

Audio annotation Guide: audio annotation

Calibration pilot

A small first batch on your own data used to agree labelling rules and edge cases before a project scales.

Start a pilot