AI & Data Annotation Glossary

A machine learning approach where the model identifies the most informative unlabeled data points and requests human annotation for those specific samples. Active learning reduces the total volume of annotations needed to achieve target model performance by focusing human effort on the data that matters most. Centric Labs uses active learning in our platform to optimize annotation efficiency.

A rectangular box drawn around an object in an image or video frame to indicate its location. Bounding boxes are defined by their top-left and bottom-right coordinates. They are the most common annotation type for object detection tasks and are used in formats like COCO, Pascal VOC, and YOLO.

A field of artificial intelligence that enables machines to interpret and understand visual information from the world, including images and video. Computer vision applications include object detection, image classification, facial recognition, autonomous driving, and medical imaging analysis. High-quality annotated image and video data is essential for training CV models.

The process of labeling data — images, video, text, audio, or 3D point clouds — to create training datasets for machine learning models. Annotations provide the ground truth that supervised learning algorithms use to learn patterns. Quality annotation is the single most important factor in building accurate, reliable AI systems.

The process of collecting, organizing, cleaning, and managing datasets to ensure they are suitable for machine learning model training. Data curation includes removing duplicates, handling class imbalances, filtering low-quality samples, and ensuring datasets are representative of real-world conditions.

An annotation technique that combines object detection and semantic segmentation by identifying each distinct object in an image and delineating its exact pixel-level boundary. Unlike semantic segmentation which groups all pixels of a class together, instance segmentation distinguishes between individual objects of the same class (e.g., separate cars in a parking lot).

A measure of how consistently multiple annotators label the same data. High IAA indicates clear annotation guidelines and well-trained annotators. Common metrics include Cohen's Kappa and Fleiss' Kappa. Centric Labs tracks IAA in real-time through our quality dashboards to identify and resolve ambiguities early.

The placement of specific points on an object to capture its shape, pose, or structure. Commonly used for human pose estimation (e.g., marking joints like shoulders, elbows, and knees), facial landmark detection, and hand gesture recognition. Keypoints are connected by edges to form a skeleton representation.

A remote sensing technology that uses laser pulses to measure distances and create detailed 3D representations of environments. LiDAR is critical for autonomous vehicles and robotics. The resulting 3D point cloud data requires specialized annotation including cuboid labeling, semantic segmentation, and sensor fusion with camera images.

A natural language processing task that identifies and classifies named entities in text into predefined categories such as person names, organizations, locations, dates, monetary values, and medical terms. NER is foundational for information extraction, question answering, and document understanding systems.

A precise annotation method where annotators draw multi-sided shapes around objects to closely follow their irregular boundaries. Polygons provide more accurate object delineation than bounding boxes, especially for non-rectangular objects like vehicles, animals, or terrain features. Used extensively in autonomous driving and satellite imagery analysis.

A training methodology used to align large language models (LLMs) with human preferences and values. Human evaluators compare pairs of model outputs and indicate which response is better, safer, or more helpful. These preference signals are used to train a reward model that guides the LLM toward producing more aligned responses. RLHF is the process behind the alignment of models like GPT-4 and Claude.

A pixel-level annotation technique that assigns a class label to every pixel in an image. Unlike object detection which locates objects, semantic segmentation provides a complete understanding of the scene by classifying every pixel as road, sidewalk, vehicle, pedestrian, sky, building, etc. Essential for autonomous driving perception and medical image analysis.

The process of combining data from multiple sensor modalities — typically cameras, LiDAR, and radar — to create a unified, more accurate understanding of the environment. In annotation, sensor fusion involves aligning and labeling objects across 2D camera images and 3D point clouds simultaneously. Critical for autonomous vehicle perception systems.

Artificially generated data that mimics the statistical properties of real-world data. Synthetic data is used to augment training datasets, address class imbalances, generate edge cases, and create training data for scenarios that are rare, dangerous, or expensive to capture in reality. Often used alongside real annotated data for optimal model performance.

The process of assigning predefined categories or labels to text documents. Common applications include sentiment analysis (positive/negative/neutral), topic categorization, intent detection in chatbots, spam filtering, and content moderation. Text classification requires annotators to read and understand context to apply the correct label consistently.

The labeled dataset used to teach machine learning models to recognize patterns and make predictions. Training data quality directly determines model accuracy — the "garbage in, garbage out" principle. Enterprise AI projects require large volumes of consistently annotated, domain-specific training data that reflects real-world conditions the model will encounter in production.