Label Assist (YOLO)

Label Assist runs an object detector over project images and writes matching boxes as annotations (source=yolo).

Requirements

pip install "oneopen-ml-studio[yolo]"

You also need at least one project class. Detections are kept only when the model class name matches a project class name (case-insensitive). Others are counted in skipped_unmatched.

How to run (UI)

  1. Open the project or annotator.

  2. Choose a model (Ultralytics catalog or a registered custom weight).

  3. Set confidence / IoU thresholds.

  4. Optionally limit to selected images.

  5. Run Label Assist, then review and fix boxes manually.

Behavior details

  • Prior annotations with source="yolo" on a processed image are removed before new ones are written.

  • Manual / SAM / import annotations are not cleared by Label Assist.

  • LibreYOLO entries appear in the train catalog; Label Assist uses Ultralytics + custom registered weights.

API

POST /api/projects/{project_id}/auto-annotate/yolo
Content-Type: application/json

{
  "model": "yolov8n.pt",
  "confidence": 0.25,
  "iou": 0.45,
  "image_ids": null
}

Response shape:

{
  "processed": 12,
  "annotations_created": 40,
  "skipped_unmatched": 3,
  "model": "yolov8n.pt",
  "classes": ["person", "car"]
}

Choosing a model

Option

Notes

YOLOv8 / v9 / v10, YOLO11, YOLO26

Downloaded by Ultralytics on first use

Custom custom:{id}

Registered project/global weights

Small models (n/s)

Faster iteration for labeling

Larger models

Better recall; slower on CPU

See Models for registration and discovery.