ultimate-ai-labeller 0.1.12

AI Labelling tool for computer vision

Ultimate AI Labeller

Desktop image annotation tool for object detection datasets (Tkinter + Ultralytics), with a separate web project.

Features

• Bounding-box annotation with drag, move, and resize handles
• Rotated bounding boxes:
  • Drag rotate knob on selected box
  • 8 resize handles follow box rotation
  • Keyboard rotate (Q/E, Shift+Q/E)
• Multi-select boxes (Shift/Ctrl + Click) for batch class reassignment and delete
• Select all boxes in current image (Ctrl+A)
• Nested/overlapping box picking prefers inner (smaller) box for easier adjustment
• Undo/redo history (Ctrl+Z, Ctrl+Y)
• Image navigation (F next/save, D previous)
• Auto red-region proposal (A)
• YOLO detection from UI (Run Detection)
• Startup source selection:
  • Open Images Folder
  • Open YOLO Dataset
  • Open RF-DETR Dataset
• Detection model management:
  • Official model mode (yolo26m.pt path by default)
  • Import custom models (.pt, .onnx) via Browse Model
  • Select model from dropdown library
• Train from existing labels:
  • Choose training range by index
  • Choose weight source before training:
    • Official yolo26m.pt
    • Custom weight file (.pt / .onnx)
    • From scratch (pretrained=False)
  • Save training artifacts to selected output folder
  • Non-blocking background training (continue labeling while training)
  • Built-in training monitor (command/log/progress/ETA)
• Class management:
  • Add / rename / delete class in class table
  • Deleting a class reindexes following class IDs automatically
• Auto-detect and propagate options
• Scrollable right settings panel
• Remove/restore bad frames from split
• Image dropdown jump
• Session resume (last project/split/image/model settings)
• English/Chinese UI switch and light/dark theme
• Export all annotations by format:
  • YOLO (.txt) export to images/train + labels/train structure
  • JSON full dataset export (per-image annotation json)

Repositories

• Desktop app (this repo): https://github.com/JamesChang666/ultimate_ai_labeller
• Web app (separate repo): https://github.com/JamesChang666/labeller_web

Dataset Structure

your_project/
  images/
    train/
    val/
    test/
  labels/
    train/
    val/
    test/

• Image extensions: .png, .jpg, .jpeg
• Label format: YOLO txt (class cx cy w h, normalized)
• Rotated-box metadata (when used): sidecar *.txt.rot.json with angles_deg
• Full guide (ZH): docs/dataset-structure-guide.md

Removed frames are moved to:

your_project/
  removed/
    train|val|test/
      images/
      labels/

Install

From PyPI:

pip install ultimate_ai_labeller

From local wheel:

pip install dist/ultimate_ai_labeller-*.whl

From source:

pip install .

For development:

pip install -e .

Run

ai-labeller

Or:

python src/ai_labeller/main.py

Web Version

This desktop repository includes local development files under web_labeller/, but the maintained web repository is:

• https://github.com/JamesChang666/labeller_web

Run locally:

cd web_labeller
pip install -r requirements.txt
uvicorn app:app --host 127.0.0.1 --port 8000 --reload

Shortcuts

• F: save and next image
• D: previous image
• A: auto red detection
• Q/E: rotate selected box (-5° / +5°)
• Shift+Q/E: rotate selected box faster (-15° / +15°)
• Ctrl+Z: undo
• Ctrl+Y: redo
• Ctrl+A: select all boxes in current image
• Delete: delete selected box

Notes

• Default detection model mode is Official YOLO26m.pt (Bundled).
• If the official model file is unavailable locally, import a custom .pt/.onnx model from the UI.
• If CUDA/GPU runtime is incompatible, detection/training automatically falls back to CPU.
• To use your own Tk app icon, put app_icon.png in src/ai_labeller/assets/.
• Session file: ~/.ai_labeller_session.json.
• Project progress YAML: <project_root>/.ai_labeller_progress.yaml (resume split/image and class names after reopen).