annobel 0.0.4

Automatic + manual YOLO bounding box annotation tool (GUI + console)

annobel

Automatic + manual YOLO-format bounding box annotation tool with a Tkinter GUI (editor) and console fallback.

Version: 0.0.3 (Alpha)

Overview

annobel streamlines creating and refining object detection datasets. It wraps Ultralytics YOLOv8 models for optional automatic bounding box proposal, then lets you refine or create labels manually. The tool stores annotations in standard YOLO text format (class x_center y_center width height, normalized values).

Main Features

• Automatic annotation (detection) using Ultralytics YOLOv8 model variants or local weights.
• Manual drawing, moving, resizing, deleting bounding boxes.
• Class list management (add / rename) saved to a classes.txt file.
• Autosave after every modification.
• Console fallback when tkinter is not present.
• Per-image navigation and keyboard shortcuts.
• Optional subset filtering of detection classes.
• Non-destructive: detection stage creates / overwrites only label files.

Installation

pip install annobel

Dependencies

• Python >= 3.8
• ultralytics (YOLO) – installed automatically.
• pillow for image handling.

Quick Start (CLI)

After installation:

annobel

You will be prompted (GUI or console fallback) to choose:

1. Mode: Automatic Annotation (YOLO) or Manual Annotation / Editing.
2. Image & label directories.
3. (Auto mode) model selection (download variant or local path) and optional class subset.

Labels are saved as <image_stem>.txt in the labels directory. A classes.txt is created/updated as needed.

Programmatic Usage

from annobel import run  

run(
    images_dir="/path/to/images",
    mode="auto",                 # or "manual"
    model_path="yolov8n.pt",      # only needed for auto mode
    conf=0.25,
    write_empty_detection_files=True,  # 0.0.3 default now ensures 1:1 images:labels
    classes_filter_ids=[0, 2],    # optional subset
)

For full interactive loop (menu reappears after exiting editor):

from annobel import main
main()

Directory Structure (Generated)

images/
  img001.jpg
  img002.jpg
labels/
  img001.txt
  img002.txt
  classes.txt

YOLO Label Format Recap

Each line: class_id x_center y_center width height with all coordinates normalized to [0,1] relative to image width/height.

Keyboard & Mouse Shortcuts (Editor)

• A: Add mode
• E: Edit mode
• C: Change class of selected box
• M: Manage classes dialog
• Delete: Remove selected box
• ← / →: Previous / next image
• Q: Return to main menu
• I: Print info about current boxes
• Mouse (Add mode): Click-drag to draw box
• Mouse (Edit mode): Drag box interior to move; handles to resize; right-click to delete

Recommended Workflow

1. Place raw images into an images directory.
2. Run auto mode with a YOLO model (e.g., yolov8n.pt) to bootstrap labels.
3. Inspect & refine with manual editor (resize, delete, add boxes, change classes).
4. Iterate until dataset is satisfactory.

Ultralytics Notice & Licensing

This package depends on the external ultralytics project for YOLO model loading and inference. Key points:

• Ultralytics repository: https://github.com/ultralytics/ultralytics
• Their code / weights are subject to the AGPL-3.0 license by default and/or alternative commercial terms offered by Ultralytics. See their LICENSE and documentation for the authoritative terms.
• This package does NOT bundle or redistribute model weights; it only invokes the YOLO API when you request detection.
• If you deploy or integrate this in a network service or commercial product, ensure you meet the obligations of the Ultralytics license (e.g., providing source, offering access, or acquiring a commercial license if required).

You (the end user) are solely responsible for verifying that your intended use of the models and resulting annotations complies with applicable licenses and data privacy regulations.

Contributing

1. Fork repository
2. Create a feature branch: git checkout -b feature/awesome
3. Install dev extras: pip install -e .[dev]
4. Run linters / tests: flake8, black --check ., pytest
5. Submit PR

Versioning / Changelog

• 0.0.3: Create empty label .txt for images with zero detections (default on). Added write_empty_detection_files=True default. Run progress line now shows whether empty label writing is ON. This ensures a strict 1:1 image:label mapping and prevents silent negatives being skipped. To restore previous behavior pass write_empty_detection_files=False to run().
• 0.0.2: Initial packaged release (renamed to annobel)

Upgrading Notes (0.0.2 -> 0.0.3)

If you previously generated datasets with missing label files for negative images, you can retroactively create empty files:

python - <<'PY'
from pathlib import Path
images=Path('path/to/images')
labels=Path('path/to/labels')
labels.mkdir(exist_ok=True, parents=True)
img_exts={'.jpg','.jpeg','.png','.bmp','.tif','.tiff','.webp'}
existing={p.stem for p in labels.glob('*.txt') if p.name!='classes.txt'}
created=0
for img in images.iterdir():
    if img.is_file() and img.suffix.lower() in img_exts and img.stem not in existing:
        (labels/f"{img.stem}.txt").write_text('')
        created+=1
print('Created empty label files:', created)
PY

Git Version Management

Recommended workflow to preserve 0.0.2 and introduce 0.0.3:

1. Commit current 0.0.2 state (if not already):

   git add . && git commit -m "chore: finalize 0.0.2 release"
   git tag -a v0.0.2 -m "annobel 0.0.2"
   

2. Apply the 0.0.3 changes (already done in this repo), then:

   git add . && git commit -m "feat: 0.0.3 default empty label file support"
   git tag -a v0.0.3 -m "annobel 0.0.3"
   git push origin main --tags
   

3. To create a maintenance branch for 0.0.2 (optional):

   git branch release/0.0.2 v0.0.2
   git push origin release/0.0.2
   

4. If you need a hotfix for 0.0.2 later: branch from release/0.0.2, bump version (e.g. 0.0.2.post1), tag, and publish.

To install a specific version for reproducibility:

pip install annobel==0.0.2   # old behavior (no auto empty files)
pip install annobel==0.0.3   # new behavior

Security / Privacy

Images are processed locally. No images or labels are uploaded by this package. Be cautious with sensitive imagery and model weight redistribution restrictions.

Disclaimer

Provided "AS IS" under AGPL-3.0-or-later. No warranty of fitness for a particular purpose. Always validate annotations before using for training.

License

• This package: AGPL-3.0-or-later (see LICENSE)
• Ultralytics YOLO: AGPL-3.0 (or commercial) – external dependency

---

© 2025 Sayali Dongre