The YOLOv8-Face repository provides pre-trained models designed specifically for face detection. The models have been pre-trained by Lindevs from scratch.

The models have been trained on WIDERFace dataset using NVIDIA RTX 4090. YOLOv8 models were used as initial weights for training.

• Training results:

• Evaluation results on WIDERFace dataset:

• "Edit configuration" before starting a Jupyter notebook:

  • Memory: 10000 (MB), CPUs: 1, GPU: 1

• Start the notebook with the environment 'yolo8'

• Create a console in Jupyter and run the following

    bash
    cd yolov8-face
    python predictsave.py --weights weights/yolov8n-face-lindevs.pt --source data/images/bus.jpg

Now check in the 'results' directory for your bus png file with detected faces

Create feature groups for all training images

    1-create-feature-groups.ipynb

Create the training/test split and fine tune the pretrained model

    2-fine-tune.ipynb

Now check in the model registry for your trained model

You need to set the following environment variables

• HOPSWORKS_HOST=saab.dev-cloud.hopsworks.ai
• HOPSWORKS_PROJECT=groupX
• HOPSWORKS_API_KEY=

python run-job create [drop]
python run-job train [drop]

Exploratory data analysis of the images/bounding-boxes in feature groups.

    3-eda-fgs.ipynb

Build vector index for similarity search in a feature group (requires a GPU). Takes a few mins.

    4-similarity-search-index.ipynb

UI with gradio to do similarity search:

    5-similarity-search-gradio.ipynb

pip install -r requirements.txt

• Download WIDERFace dataset and annotations:

python download.py

• Convert annotations to YOLO format:

python annotations.py