YOLOv8-Face

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

Release Notes

• [2023-12-09] YOLOv8x-Face model has been added.
• [2023-12-02] YOLOv8n-Face, YOLOv8s-Face, YOLOv8m-Face and YOLOv8l-Face models has been added.

Pre-trained Models

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

Name	Image Size (pixels)	mAPval 50-95	Params	GFLOPs
YOLOv8n-Face	640	37.5	3005843	8.1
YOLOv8s-Face	640	40.6	11125971	28.4
YOLOv8m-Face	640	41.7	25840339	78.7
YOLOv8l-Face	640	42.8	43607379	164.8
YOLOv8x-Face	640	43.3	68124531	257.4

• Download links:

Name	Model Size (MB)	Link	SHA-256
YOLOv8n-Face	6.3 12.2	PyTorch ONNX	a72e65818b9d61b15d6b5f58996b2391c827f4508c5725706cf61d05a9a70e49 3ebbeaddfe4fc51a9ab119a4c67cc6193b559b07a8104d029ba0e146fcaa2ad4
YOLOv8s-Face	22.5 44.7	PyTorch ONNX	3c9962b02c911b0ff6e4bb5d634f6e72d0d8ba24ca3442287f280ba50bfe2f73 0bf65e2576c05f25a5d5454b57b313d54d41495b044e94eae25bb1205e5d8d18
YOLOv8m-Face	52.0 103.6	PyTorch ONNX	2d96e2eac09fcac4a677664680beee1c210041d0eb7e2f6a434fb806d455b2dc 5c40fbed7e8c2328ccaca718eb9ce49e5d631ae622c275dbd569ef4feab70ebd
YOLOv8l-Face	87.6 174.7	PyTorch ONNX	a24036da36c2b8ce5e2985f70b3f8bc0bd1df3941f48e91ec84b48bdd73345b4 5037c8362da630935b35f32670b732d4d196bfcd6eff5052704bb9a568955e7f
YOLOv8x-Face	136.7 272.7	PyTorch ONNX	1985c9b08ee443f0fff646cc0371dbff9825415fb5e33d33bd2dc14bd73a56d7 28d9e9a85e8e739f824f280b0f426cc77b0984ebac6dbc30817ae8e3336ac97e

• Training results:

Name	Training Time	Epochs	Batch Size	Non-default parameters	Link
YOLOv8n-Face	2.75 hours	300	16	-	results.txt
YOLOv8s-Face	2.68 hours	200	16	-	results.txt
YOLOv8m-Face	3.01 hours	120	16	-	results.txt
YOLOv8l-Face	3.97 hours	110	16	-	results.txt
YOLOv8x-Face	13.65 hours	240	16	optimizer='SGD' lrf=1e-5 weight_decay=5e-3	results.txt

• Evaluation results on WIDERFace dataset:

Name	Easy	Medium	Hard
YOLOv8n-Face	93.79	91.82	79.38
YOLOv8s-Face	95.13	93.62	82.90
YOLOv8m-Face	95.73	94.47	84.55
YOLOv8l-Face	96.26	95.03	85.43
YOLOv8x-Face	96.33	95.16	85.80

Instructions

Installation

pip install -r requirements.txt

Prediction

python predict.py --weights weights/yolov8n-face-lindevs.pt --source data/images/bus.jpg

• OpenCV DNN

python examples/opencv-dnn-python/main.py --weights weights/yolov8n-face-lindevs.onnx --source data/images/bus.jpg

Export

• Install package:

pip install onnx

• Export to ONNX format:

python export.py --weights weights/yolov8n-face-lindevs.pt

• Or export to ONNX format using dynamic axis:

python export.py --weights weights/yolov8n-face-lindevs.pt --dynamic

Dataset Preparation

• Download WIDERFace dataset and annotations:

python download.py

• Convert annotations to YOLO format:

python annotations.py

• Copy widerface.yaml.example file to widerface.yaml:

python data_file.py

Training

• Prepare dataset.
• Start training:

python train.py --weights yolov8n.pt --epochs 300 2>&1 | tee -a results.txt
python train.py --weights yolov8s.pt --epochs 200 2>&1 | tee -a results.txt
python train.py --weights yolov8m.pt --epochs 120 2>&1 | tee -a results.txt
python train.py --weights yolov8l.pt --epochs 110 2>&1 | tee -a results.txt
python train.py --weights yolov8x.pt --epochs 240 --optimizer SGD --lrf 1e-5 --weight-decay 5e-3 2>&1 | tee -a results.txt

• Or resume training:

python train.py --weights runs/detect/train/weights/last.pt --resume 2>&1 | tee -a results.txt

Validation

• Prepare dataset.
• Start validation:

python validate.py --weights weights/yolov8n-face-lindevs.pt

WIDERFace Evaluation

• Prepare dataset.
• Start prediction on validation set:

python widerface/predict.py --weights weights/yolov8n-face-lindevs.pt

• Install package:

pip install Cython

• Build extension:

cd widerface && python setup.py build_ext --inplace && cd ..

• Start evaluation:

python widerface/evaluate.py