louisfacun / facemask-detection-yolov4

Wearing of Face Mask Detection Using YOLOv4

This repository contains the specific files and YOLOv4 implementation we used in our research paper.

Prerequisites

To keep this repository simple we won't include the source code of YOLOv4.
Install YOLOv4's darknet. See YOLOv4's install guide https://github.com/AlexeyAB/darknet#requirements-for-windows-linux-and-macos
If darknet (YOLOv4) is successfully installed, start running model inference using our files: facemask.weights facemask.cfg facemask.data facemask.names

How to use

The following files are specific in this study:

facemask.weights (244 MB)
facemask.cfg

facemask-detection-yolov4/facemask.cfg

Lines 4 to 5 in 86405f6

width=640

height=640
facemask.data

facemask-detection-yolov4/facemask.data

Lines 1 to 2 in a4cd233

classes = 3

names = facemask.names
facemask.names

facemask-detection-yolov4/facemask.names

Lines 1 to 3 in d8098be

NWM

WM

IWM

Files and directory structure

Put facemask.weights facemask.cfg facemask.data facemask.names in the same folder of compiled YOLOv4's darknet.

.
.
darknet (created after compiling darknet)
.
.
facemask.weights
facemask.cfg
facemask.data
facemask.names
darknet_image.py
darknet_video.py

Running an inference from different source

Single image: python darknet_image.py --input "<image location>"
PC webcam: python darknet_video.py
Video file: python darknet_video.py --input "<video location>"
IP camera: python darknet_video.py --input "<rstp ip>"

Tools and libraries we used

LabelImg: https://github.com/tzutalin/labelImg (labelling and drawing bounding boxes on images)
Albumentations: https://albumentations.ai/docs/getting_started/installation/ (applying augmentations)

Model info

Labels

Label id	Label	Description	Box Color
0	NWM	Not Wearing a Mask	Red
1	WM	Wearing a Mask	Green
2	IWM	Improperly Wearing a Mask	Yellow

Dataset stats

Label	Instances	Min Dimension	Max Dimension
NWM	1,006	14x17	235x265
WM	1,012	14x15	614x605
IWM	1,007	22x53	559x577

Results

Average inference time	mean Average Precision (PASCAL VOC 2012)
73 ms	94%

*Using GTX 1650 Super (CUDA Enabled Inference Time)

Confusion matrix

None row in the predictions means, for example, that there is 3 ground-truth of IWM that is not predicted.

Sample detections

Citation

If you use our study as a reference in a scientific publication, we would appreciate citations to the following:

Object Detection Frameworks fo Real-Time, Scale-Invariant Face Mask Detection, Facun et al., IJFCC vol. 11, no. 1, pp. 12-17, 2022.

@article{facun_facemask_2022,
    title={Object Detection Frameworks for Real-Time, Scale-Invariant Face Mask Detection},
    author={Facun, Louis Philippe B. and Baculo, Maria Jeseca C. and Libao, Marlon F.
            and Eisma, Ceazar M. and Fredeluces, Christian B. and Garlejo, Rigzor A.
            and Idio, Raymart S.},
    journal={International Journal of Future Computer and Communication},
    volume={11},
    number={1},
    pages={12-17},
    year={2022}
}