This is a forked repository from https://github.com/hunglc007/tensorflow-yolov4-tflite. While a majority of the code from between these two repository remains the same, there are a few notable changes. This repository does not contain training, evaluating, benchmarking and weight conversion functionalities from the parent repository. The goal was to create a customised version of just for real-time face mask detection using the trained darknet weights.
The parent repository allowed for several general functionalities, most of which was unnecessary for our use case. As a result this repository will only be maintained with the sole intention of working with face mask detection problem.
To see the full list of changes, please read through the Changelog.
The purpose of this project is develop a real-time face mask detection software that can be used in a wide variety of places such as, corporate offices, shopping malls, theatres and other places with dense crowds. To develop this application we have used YOLOv4 a real-time object detection model. We used the publicly available face mask detection dataset from Kaggle and trained it using darknet's pre-trained yolov4 weights. We trained on an initial network resolution of 608x608. Higher resolution can lead to better detection.
Each of the trained weights was converted to an .h5 model which was used for detection.
To run the detection on a single image :
python image.py --weights <path_to_the_weights> --size <size_to_scale> --image_path <path_to_image>Detection on video and to save the results to a separate file :
python video.py --weights <path_to_the_weights> --video_path <path_to_the_video> --size <size_to_scale> --save_path <detected_video_save_path>Detection using a webcam :
python webcam.py --weights <path_to_the_weights> --size <size_to_scale> size parameter is used to scale the image to get a padded image for the detection. Since the input_layer uses a tf.keras.layers.Input with the size as the shape of this layer, each image / video frame sent to this model for detection will be scaled accordingly.