Reference

Glossary

Definitions of the terms used across BOAT, its CLI and this documentation.

Annotation type. The kind of label a project produces: object detection, oriented object detection, instance segmentation, semantic segmentation, crowd counting, keypoint detection or classification. Each project has exactly one.

Auto label. A tool that labels the whole image without a prompt.

Auto search. A tool that takes a loose box around one object and returns a tight mask or rotated box for it.

Backup. A timestamped archive of BOAT's data created with boat backup create, and automatically before every update.

Candidate picker. An apply mode where the model returns several suggestions and you choose which to keep.

Cascading. A detection tool that treats every box you draw as a prompt and suggests similar objects automatically.

Channel. The release stream boat update follows: stable or beta.

CLI. The boat command. It installs, configures, runs, updates and backs up everything else.

Common archive. The part of a model package shared by all platforms: adapter code, tool definitions, icons and licenses.

Core. The BOAT web application: the backend API and the annotation editor.

Default Device. The device models use unless configured otherwise: auto, cpu, mps or cuda:<n>.

Editor binding. A tool's connection to an existing editor flow, such as instance-segmentation_point-prompt. Tools without one use the generic tool panel.

Gateway. The small process in front of Core that serves the web app and supervises the backend.

Idle time. How long a loaded model worker stays in memory without requests before it unloads. Defaults to 300 seconds.

Manifest. The signed JSON document that describes one release of a model package: tags, weights, runtimes, tools and license.

Model package. An installable unit that adds one model and its tools. Identified by a package ID such as sam3-litetext.

Model reference. <id> or <id>:<tag>, as used by boat model commands.

Negative prompt. A box or point marking an area the model should leave out.

Recommended score. A 0–10 rating in the registry that ranks models on this website. Higher means a better default choice for most people.

Registry. The static bucket at models.boat-annotation.com that hosts index.json, manifests and blobs.

Role. A preset of project permissions: Personal owner or collaborator, and Team owner, admin, labeler or viewer.

Runtime profile. The platform build a model runs with: macos-metal, linux-cpu, linux-cuda126 or linux-cuda128.

Tag. An installable variant of a model, such as a size or quantization. Also called a variant.

Tool. One capability a model adds to the editor, defined by its input (text, box, point or none), its output and the annotation types it supports.

VRAM required. The memory the model's weights need on the device, across its tags. BOAT reserves about 0.5 GB more when it loads a worker.

Worker. An isolated Python process that runs one installed model. Tools from the same model share its worker.

Workspace. A container for projects. Everyone has one personal workspace and can create any number of team workspaces.

Workspace model. A package, such as YOLO, that runs weights you upload to a workspace instead of downloading fixed weights.

↑↓ navigate↵ open