metadata

title: YOLOv8 Face Mask Detection
license: mit
sdk: static
colorFrom: indigo
colorTo: blue
pinned: false
description: YOLOv8 model trained on Pascal VOC annotations for face mask detection

YOLOv8 Face Mask Detection using Custom Pascal VOC Annotations

This repository presents an end-to-end pipeline for training a YOLOv8 object detection model to identify the following classes:

with_mask
without_mask
mask_weared_incorrect

The dataset used is the Face Mask Detection dataset provided by Andrew Mvd on Kaggle, originally annotated in Pascal VOC (XML) format. This implementation converts the annotations to YOLO format and trains a YOLOv8 model using the Ultralytics framework.

Features

XML annotation parsing and conversion to YOLO format
Automatic data split into training and validation sets
Directory structuring compliant with YOLOv8 training requirements
Training using YOLOv8 with configurable parameters
Evaluation and inference on the validation set
Result visualization with training metrics

Dataset

Dataset: Face Mask Detection by Andrew Mvd

Annotation format: Pascal VOC (XML)

Environment Setup

pip install ultralytics xmltodict matplotlib pandas

Directory Structure

working/
├── images/
│   ├── train/
│   └── val/
├── labels/
│   ├── train/
│   └── val/
├── data.yaml
├── face-mask-yolov8/
│   └── results.csv
└── yolov8n-mask-detection.pt

Training

from ultralytics import YOLO

model = YOLO("yolov8n.pt")  # Use yolov8s.pt or larger for improved accuracy

model.train(
    data="/kaggle/working/data.yaml",
    epochs=25,
    imgsz=640,
    batch=16,
    project="/kaggle/working",
    name="face-mask-yolov8",
    exist_ok=True
)

Evaluation and Inference

metrics = model.val()

results = model.predict(
    source="/kaggle/working/images/val",
    save=True,
    conf=0.4
)

model.save("yolov8n-mask-detection.pt")

Visualization of Training Metrics

import pandas as pd
import matplotlib.pyplot as plt

results_dir = "/kaggle/working/face-mask-yolov8"
df = pd.read_csv(f"{results_dir}/results.csv")

df.plot(
    x="epoch",
    y=["train/box_loss", "train/cls_loss", "metrics/mAP_0.5"],
    figsize=(12, 6),
    title="YOLOv8 Training Metrics"
)
plt.grid(True)
plt.show()

Inference Output

The model outputs prediction images with bounding boxes drawn and saves them in the specified directory. Results can be found in:

/kaggle/working/runs/detect/predict

License

This project is released under the MIT License.