Skip to main content
Definitions of key terms used throughout the Avala documentation.

A

Annotation

A label or markup applied to a dataset item, such as a bounding box around an object in an image or a classification tag. Annotations are the primary output of labeling projects.

Annotation Type

The kind of annotation applied to data. Avala supports several annotation types including bounding boxes, polygons, polylines, keypoints, cuboids, segmentation masks, and classification labels.

API Key

A secret token used to authenticate requests to the Avala API. API keys can be scoped to limit permissions. Manage keys in Mission Control under Settings.

Auto-labeling

The process of automatically generating annotations using machine learning models. Auto-labeling can be used to pre-annotate data before human review, significantly reducing annotation time.

B

Bounding Box

A rectangular annotation defined by its position and dimensions, used to identify and locate objects in images or video frames.

C

Channel

A named stream of messages within a dataset. Each channel has a topic name and a schema defining the message format.

Chunk

In MCAP files, a compressed block of messages. Chunks enable efficient random access and parallel processing.

Consensus

A quality control mechanism where multiple annotators label the same data item independently. Disagreements between annotators are flagged for review, improving annotation accuracy.

Cuboid

A three-dimensional bounding box annotation used to identify objects in 3D space, commonly used for point cloud and multi-sensor data.

D

Dataset

The fundamental unit of organization in Avala. A dataset contains data items such as images, videos, point clouds, or sensor recordings, along with metadata and annotations.

Dataset Item

A single piece of data within a dataset, such as an individual image, video frame, or point cloud scan. Each item can have annotations and metadata attached to it.

E

Export

The process of downloading annotations and data from Avala in a structured format (such as COCO, YOLO, or Pascal VOC) for use in model training or other workflows.

F

Frame

A single image or time step within a video or sequence. In video annotation projects, frames are the individual units that receive annotations.

I

Ingest

The process of uploading data to Avala, including validation, indexing, and processing.

K

Keypoint

A point-based annotation that marks specific locations on an object, often used for pose estimation. Keypoints can be connected to form a skeleton structure.

L

Label

A named category applied to an annotation. Labels are defined in a project’s taxonomy and represent the classes that annotators assign to objects (e.g., “car”, “pedestrian”, “stop sign”).

Log Time

The timestamp when a message was recorded, typically from the system clock at recording time.

M

MCAP

An open-source container file format for multimodal log data. MCAP supports multiple serialization formats and provides efficient random access.

MCP (Model Context Protocol)

An open protocol that allows AI assistants like Claude, Cursor, and VS Code Copilot to interact with external tools and services. Avala provides an MCP server that enables AI agents to query datasets, projects, and annotations.

Message

A single data record within a channel, containing serialized sensor data and timestamps.

Mission Control

Avala’s web-based platform for managing, visualizing, and collaborating on autonomous vehicle data and annotation projects.

P

Pipeline

An automated workflow for ingesting data from external sources into Avala.

Polygon

A closed-shape annotation defined by a series of vertices, used to outline irregularly shaped objects with more precision than bounding boxes.

Polyline

An open-shape annotation defined by a series of connected points, commonly used to annotate lanes, roads, or other linear features.

Protobuf

Protocol Buffers, a language-neutral serialization format developed by Google. One of several message encodings supported by MCAP.

Publish Time

The original timestamp when a message was published by its source, which may differ from log time.

Q

Quality Control

The set of processes and tools used to ensure annotation accuracy and consistency. Avala supports quality control through consensus workflows, review stages, and acceptance criteria.

R

Review

A quality assurance step where completed annotations are inspected by a reviewer. Reviewers can accept, reject, or request corrections to annotations before they are finalized.

ROS Bag

A file format for storing ROS (Robot Operating System) message data. Avala supports both ROS 1 and ROS 2 bag formats.

S

Schema

A definition of a message structure. Schemas can be defined using ROS message definitions, Protobuf, JSON Schema, or other formats.

Segmentation

An annotation type that assigns a class label to every pixel in an image, producing a dense pixel-level mask. Used for tasks like scene understanding and semantic segmentation.

Sequence

An ordered group of dataset items that represent a continuous recording or related set of data, such as consecutive video frames or a lidar scan series.

Sequence Status

The processing state of a sequence within a dataset. Possible statuses include pending, processing, ready, and failed.

Slice

A filtered subset of a dataset created by applying criteria such as metadata values, labels, or annotation properties. Slices are useful for organizing data for targeted annotation or model evaluation.

Splat

A 3D Gaussian splat representation used for novel view synthesis and 3D scene visualization in Mission Control.

T

Task

A unit of work assigned to an annotator within a project. Each task typically corresponds to one or more dataset items that need to be annotated or reviewed.

Task Status

The current state of a task in the annotation workflow. Common statuses include queued, in progress, completed, and rejected.

Topic

A named endpoint for publishing or subscribing to messages. In Avala, topics are represented as channel names (e.g., /camera/front).

Tracking

An annotation method used in video and sequential data where objects are tracked across multiple frames. Tracking annotations maintain a consistent identity for each object over time.

V

Vehicle

A unique identifier for an autonomous vehicle in your fleet. Datasets can be associated with specific vehicles for organization and analysis.

W

Work Batch

A group of tasks assigned together to an annotator or a team. Work batches help manage the distribution and scheduling of annotation work across a project.