Things don't always pan out the way we hope, so I couldn't add anything noteworthy to the repo lately. PhD projects are taking up all my time right now, but no stress, I'm working on something big and exciting! Keep an eye out 👀
I took small break due to other projects related to my PhD however I plan to update models before 2024. I want to try providing also 68 2D facial keypoints to obtain.
YOLOv8 for Face Detection. The project is a fork over ultralytics repo. They made a simple interface for training and run inference. Model detects faces on images and returns bounding boxes, score and class.
The next table presents the performance of the different model on some hardware configurations and their evaluation over the different subsets (easy, medium, hard) of the WiderFace validation set.
We added results from multi-scale testing using NMS and Weighted Boxes Fusion from ZFTurbo/Weighted-Boxes-Fusion repo to obtain the final bounding boxes. We used original scales and [500, 800, 1100, 1400, 1700] as those were the ones used in RetinaFace (we still lack of flip).
| Model | YOLOv8 nano | YOLOv8 medium | YOLOv8 large | RetinaFace-MNet0.251 | RetinaFace-R501 |
|---|---|---|---|---|---|
| Avg. FPS (RTX 4090) | 169 | 117 | --- | 44 | 39 |
| Avg. FPS (Colab Tesla T4) | 82 | 31 | --- | 25 | 20 |
| Avg. FPS (GTX 1650 with Max-Q Design) | 55 | 28 | --- | 19 | 16 |
| WiderFace Easy Val. AP | 0.8833 | 0.9759 | 0.9741 | 0.8382 | 0.9067 |
| +Multi-Scale+NMS | 0.8835 | 0.9740 | 0.9699 | --- | --- |
| +Multi-Scale+Weighted Boxes Fusion | 0.8902 | 0.9747 | 0.9713 | --- | --- |
| WiderFace Medium Val. AP | 0.8273 | 0.9483 | 0.9518 | 0.7678 | 0.8602 |
| +Multi-Scale+NMS | 0.8324 | 0.9481 | 0.9492 | --- | --- |
| +Multi-Scale+Weighted Boxes Fusion | 0.8439 | 0.9499 | 0.9511 | --- | --- |
| WiderFace Hard Val. AP | 0.6049 | 0.7695 | 0.7936 | 0.4320 | 0.5520 |
| +Multi-Scale+NMS | 0.6356 | 0.7864 | 0.8072 | --- | --- |
| +Multi-Scale+Weighted Boxes Fusion | 0.6416 | 0.7923 | 0.8156 | --- | --- |
I recommend to use a new virtual environment and install the following packages:
# Install pytorch for CUDA 11.7 from pip
pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117or alternatively if you are using conda
# Install pytorch for CUDA 11.7 from conda
conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidiathen you install the ultralytics package (this code works for ultralytics-8.0.21) and the ensemble boxes package.
pip install ultralytics
pip install ensemble-boxesThere are a few python scripts, train.py is to train a yolov8 model, test.py is to test the model with images and demo.py is to launch a real-time demo of the model with your webcam.
You must configure wider.yaml according to the path in your pc (default settings are relative to datasets folder).
You must download the WIDERFACE official evaluation code and results and extract the ground_truth folder that contains easy, medium and hard validation subsets.
Then you must run python evaluation.py -w <pretrained_model> -p <new prediction path> -e <path val images> to create .txt evaluation files in the proper format.
Finally you run cd Evaluation, build Cython code with python setup.py build_ext --inplace and evaluate your .txt files with python evaluation.py -p <your prediction dir> -g <groud truth dir> to get Val. AP per each subset.
WiderFace-Evaluation code is extracted from wondervictor/WiderFace-Evaluation repo.
The training arguments
parser = argparse.ArgumentParser(description='Train YOLOv8 for Face Detection')
parser.add_argument('-b', '--batch', type=int, help='Batch Size', default=8)
parser.add_argument('-e', '--epochs', type=int, help='Number of Epochs', default=100)
parser.add_argument('-w', '--workers', type=int, help='Workers for datalaoder', default=2)
parser.add_argument('--pretrained', action='store_true', help='Finetune')
parser.add_argument('--no-pretrained', dest='pretrained', action='store_false', help='Train from zero.')
parser.set_defaults(pretrained=False)
parser.add_argument('-i', '--imgsize', type=int, help='Image size', default=640)
parser.add_argument('-m', '--model',
type=str,
choices=['yolov8n','yolov8s', 'yolov8m', 'yolov8l', 'yolov8xl'],
default='yolov8m',
help='YOLOv8 size model')the test arguments
parser.add_argument('-w', '--weights', type=str, help='Path to trained weights', default='runs/detect/train/weights/best.pt')
parser.add_argument('-t', '--threshold', type=float, help='Score threshold',
default=0.5)
parser.add_argument('-i', '--input', type=str, help='Sample input image path',
default='test_input.jpg')
parser.add_argument('-o', '--output', type=str, help='Output image path',
default='test_output.jpg')evaluation arguments
parser.add_argument('-e', '--eval', type=str, help='Path to WIDER_val/images',
default=os.path.join('WIDER_val','images'))
parser.add_argument('-w', '--weights', type=str, help='Path to trained weights',
default='yolov8m_200e.pt')
parser.add_argument('-p', '--pred', type=str, help='Path to create evaluation .txt s',
default='WIDER_pred')
parser.add_argument('-t', '--threshold', type=float, help='Score threshold',
default=0.0001)
parser.add_argument('--multi-scale', action='store_true', help='Multi scale testing', default=False)
parser.add_argument('--vote', type=str, choices=['fusion','nms'], default='fusion', help='Bounding Box vote method')and finally the demo arguments
parser.add_argument('-w', '--weights', type=str, help='Path to trained weights', default='runs/detect/train/weights/best.pt')
You can download the pretrained weights (YOLOv8 large) v0.1 for Face Detection here.
You can download the pretrained weights (YOLOv8 medium) v0.2 for Face Detection here.
You can download the pretrained weights (YOLOv8 nano) v0.1 for Face Detection here.
You can also download the WiderFace dataset properly formatted to train your own model here.
Footnotes
-
RetinaFace based on hphuongdhsp/retinaface repo that is built on top of the biubug6/Pytorch_Retinaface implementation.
R50 = ResNet-50 and MNet0.25 = MobileNet-0.25 ↩ ↩2