Annotation

Image (CV) annotation lives at /projects/{id}/annotate (annotate.html). Non-image modalities use the same URL with annotate_samples.html (text, tabular, audio, video, LLM).

Computer vision (visual)

Prerequisites

At least one class on the project.
At least one image (upload or import).

Project types

Type	Label tools
`object_detection`	Bounding boxes
`instance_segmentation` / `semantic_segmentation`	Polygons (SAM can produce polygons from point prompts)
`classification`	Whole-image class assignment

There is no pixel-mask brush yet — segmentation is polygon-based. Keypoints, OBB, OCR, and point cloud are not shipped.

Tools

Tool	Geometry	Typical use
Bounding box	`{x, y, width, height}` normalized	Object detection
Polygon	Point list normalized	Instance outlines / SAM refine
Classification	Whole-image class id	Image classification

Coordinates are normalized to the image (0–1), so labels stay valid if you resize later during version generation.

Workflow tips

Select a class, then draw (or assign class for classification projects).
Edit or delete existing annotations as needed.
Use keyboard / UI navigation to move between images.
Delete the current image from the annotator when cleaning the set.
Re-run Label Assist carefully — YOLO assist clears previous source=yolo annotations on each image it processes before writing new ones. Manual annotations are kept.
Optional: SAM point prompts → polygons (see SAM (interactive segmentation)); YOLO Label Assist (see Label Assist (YOLO)).

Annotation sources

Source	Meaning
`manual`	Drawn in the UI
`yolo`	Created by Label Assist
`sam`	Created by SAM point prompts
`import`	Loaded from YOLO labels on import

CV annotation API

GET    /api/images/{image_id}/annotations
POST   /api/images/{image_id}/annotations
PUT    /api/annotations/{annotation_id}
DELETE /api/annotations/{annotation_id}

Example create body:

{
  "class_id": 1,
  "annotation_type": "bbox",
  "geometry": {"x": 0.1, "y": 0.2, "width": 0.3, "height": 0.4},
  "confidence": null,
  "source": "manual"
}

annotation_type: bbox, polygon, classification (mask exists in the schema but is unused in the UI).

Upload images

UI: /projects/{id}/upload — multi-file upload.

API:

POST /api/projects/{project_id}/images
Content-Type: multipart/form-data
files: <one or more images>

Supported extensions: .jpg, .jpeg, .png, .bmp, .webp, .tif, .tiff.

Files are stored under {project_location}/images/ with UUID filenames.

Delete images

Single: DELETE /api/images/{image_id}
Bulk: select mode on the project grid → POST /api/images/bulk-delete

Non-image modalities

For text, tabular, audio, video, and LLM projects:

Create the matching project type.
Import CSV/JSONL (see Text and tabular modalities and Audio, video, and LLM datasets).
Open Annotate — the sample UI labels spans, values, transcripts, or LLM fields.
Use sample APIs under /api/projects/{id}/samples and /api/samples/{id}.

In-process training is deepest for CV (+ sklearn for text/tabular). Audio/video/LLM emphasize curate → package for external trainers.