Mask Detection System

Mask Detection System using Yolov3 on custom made dataset : https://github.com/archie9211/Mask-Detection-Dataset

This project uses Yolov3 for the mask detection on live stream video from either webcam or cctv.

Dataset Used : Repo Link

Sample Output on intial training :

• Yolo weights given below were trained on a imbalanced dataset (5000 + mask samples and < 1000 non mask samples.). So, Its performance can be poor. This model was only trained for 20K batches of batch size 64.

• Other Yolov3 models are typically trained for more than 500K batches.

I am training the model on new dataset handicked and annotated using https://github.com/tzutalin/labelImg.

which has 4k images of each category. So, Improvement is expected. : link

Usage

clone the repo

https://github.com/archie9211/Mask-detection-system
cd Mask-detection-system 
pip3 install -r requirements.txt

then download the yolo weights

wget https://my.gdrivemirror.workers.dev/yolov3_final.weights -O data/yolov3.weights
python convert.py

Running simple yolo on a video on webcam (--video 0)

there are three modes of execution :

python run.py --video /path/to/video 

you can use another flag for saving the output to a video

--video /path/to/video.mp4

run following command for full help of arguments and their uses :

python run.py --help

for training:

this will need voc format dataset for training.

Example command line arguments for training

python train.py --batch_size 8 --dataset ~/Data/voc2012.tfrecord --val_dataset ~/Data/voc2012_val.tfrecord --epochs 100 --mode eager_tf --transfer fine_tune

python train.py --batch_size 8 --dataset ~/Data/voc2012.tfrecord --val_dataset ~/Data/voc2012_val.tfrecord --epochs 100 --mode fit --transfer none

python train.py --batch_size 8 --dataset ~/Data/voc2012.tfrecord --val_dataset ~/Data/voc2012_val.tfrecord --epochs 100 --mode fit --transfer no_output

python train.py --batch_size 8 --dataset ~/Data/voc2012.tfrecord --val_dataset ~/Data/voc2012_val.tfrecord --ep

for yolo format dataset training : go to darknet folder

cd darknet

and make sure dataset is in imgs folder

folder structure of darknet folder should be same as https://github.com/archie9211/Mask-Detection-Dataset

there is additional compiled binary of darknet

download pretrained weights for training

wget https://my.gdrivemirror.workers.dev/darknet53.conv.74

for training

./darknet detector train "obj.data" "yolov3.cfg" "darknet53.conv.74"  -dont_show

you can also train over my weights

./darknet detector train "obj.data" "yolov3.cfg" "../data/yolov3.weights"  -dont_show

files to mentions :

 train.py
 yolov3-tf3/models.py
 run.py
 detect.py

References

https://github.com/zzh8829/yolov3-tf2 (yolo implementation)

https://medium.com/@artinte7/real-time-object-detection-using-yolo-upon-google-colab-in-5-minutes-fd65a4903df5