Using BOAT
Annotation types and tools
The seven annotation types, the model tools each one supports, and how the editor uses them.
Each project has one annotation type. The editor for that type offers manual tools, plus AI tools from the models you have installed and enabled. Tools that do not fit the current annotation type are hidden, and uninstalled or disabled tools never appear.
The seven types
| Type | You produce | Typical use |
|---|---|---|
| Object detection | Axis-aligned bounding boxes | Vehicles, products, defects |
| Oriented object detection | Rotated boxes | Aerial imagery, ships, text lines |
| Instance segmentation | One mask per object | Counting and measuring individual objects |
| Semantic segmentation | A class for every pixel | Road scenes, land cover, medical regions |
| Crowd counting | One point per object | Wildlife, crowds, cells |
| Keypoint detection | Skeleton keypoints | Human pose |
| Classification | One label per image | Quality control, disease classes |
AI tools
Models plug into the editor through a small set of standard tools. When several installed models provide the same tool, the tool shows one button with a Model picker.
| Tool | Annotation types | You give | You get |
|---|---|---|---|
| Text prompt | Detection, instance and semantic segmentation, crowd counting | A class name such as forklift |
Boxes, masks or points for every match |
| Box prompt | Detection | One example box | Boxes for similar objects, as candidates |
| Box prompt | Instance and semantic segmentation | One or more boxes, optionally negative | Masks for exactly those objects |
| Point prompt | Instance and semantic segmentation | Positive and negative clicks | A precise mask |
| Auto search | Instance and semantic segmentation, oriented detection | A loose box around an object | A tight mask or rotated box |
| Cascading | Detection | Nothing extra; every box you draw becomes a prompt | Suggested boxes for look-alike objects |
| Auto label | Detection, keypoint detection | Nothing | Annotations for the whole image |
Each model page shows exactly which tools that model provides, with a looping example.
Previews and candidates
- Preview tools draw results on the canvas first. Apply them to save, or adjust the prompt and run again.
- Candidate picker tools return several suggestions. Accept the ones you want and discard the rest.
Applied results go through the editor's normal save and undo history, so AI output is edited exactly like manual annotations.
Sessions
Changing the model, model version, image or class ends the current prompt session. Editing a tool parameter keeps your prompts but cancels any request in flight and clears earlier candidates.
Workspace models
Some packages, such as YOLO, run weights that you upload to a workspace instead of downloading fixed weights. They provide the Auto label tool for detection and keypoint projects. Uploaded weights are validated with a real forward pass and stay in your workspace even if the package is removed, then work again after you reinstall it.
Classification
The classification editor is designed for fast keyboard labeling and currently has no model tools.