AutoLabeller

An intelligent dataset labeling tool combining YOLO-World's open-set detection capabilities with human verification, optimized for quickly creating high-quality COCO-format datasets.

Features

• YOLO-World Integration: Leverage state-of-the-art open-set object detection
• Smart Frame Selection: Avoid redundant frames using detection-based similarity checks
• Interactive GUI:
  • Real-time bounding box editing
  • Class cycling with single click
  • Drag-to-draw missing detections
  • Dynamic confidence threshold adjustment
• COCO Dataset Export: Industry-standard output format
• Performance Optimizations:
  • GPU-accelerated inference
  • Multi-threaded processing
  • Intelligent frame caching

Pipeline Overview

graph TD;
A[Input Video/Images] --> B[YOLO-World Detection]
B --> C[Frame Filtering]
C --> D[Manual Verification GUI]
D --> E[COCO Dataset Export]
E --> F[Training Ready Dataset]

Loading

Inspiration

This project was born from the need to:

1. Reduce manual labeling effort in custom detection projects
2. Create a middle-ground between fully automatic and manual labeling
3. Leverage YOLO-World's exceptional open-vocabulary capabilities
4. Provide a standardized COCO-format output for compatibility with modern ML frameworks

Installation

Requirements

• Python 3.8+
• NVIDIA GPU (recommended)
• CUDA 11.8
• cuDNN 8.6+

From Source

git clone https://github.com/sah4jpatel/AutoLabeller.git
cd AutoLabeller
pip install -e .

Usage

Basic Command

autolabel
--input path/to/input
--classes "class1" "class2" ...
--model path/to/model.pt
--output dataset.json

Full Options

autolabel
--input path/to/video_or_directory
--classes class1 class2 class3
--model custom_yoloworld.pt
--output my_dataset.json
--conf_thresh 0.15
--full_video
--min_images 1000

Flag	Description	Default
--input	Video file or image directory	Required
--classes	Space-separated list of classes	Required
--model	YOLO-World model path	Required
--output	Output COCO JSON path	dataset.json
--conf_thresh	Detection confidence threshold	0.25
--full_video	Disable frame filtering	False
--min_images	Minimum output images	0

Training Custom YOLO-World Models

1. Prepare dataset in COCO format
2. Create model config (custom.yaml):

path: datasets/custom
train: images/train
val: images/val
names:
    0: class1
    1: class2

3. Start training:

yolo detect train data=custom.yaml model=yolow_world.pt epochs=100 imgsz=640 device=0

Full training documentation: Ultralytics YOLO-World Docs

Troubleshooting

Q: Getting ModuleNotFoundError: No module named 'autolabeller'
A: Ensure proper installation using pip install -e . from project root

Q: GUI window scaling issues
A: Install screeninfo package: pip install screeninfo

Q: CUDA out of memory errors
A: Reduce batch size in detection parameters or use smaller model variant

License

MIT License - see LICENSE for details

---