Multithreading is used to speed up face detection using Tensorflow MTCNN model
Observed > 40FPS for 1920 x 1080 webcam feed on 16GB CPU RAM and intel i7 processor
- Tensorflow > 1.0
- Fork and install Tensorflow Object Detection API: https://github.com/tensorflow/models/tree/master/research/object_detection
- OpenCV using
pip install opencv-contrib-python - Fork the new trained model (
20180408-102900) for Face Detection and Recognition from https://github.com/arunmandal53/facematch
In this script the default video source is set to webcam with width and height as 1920 x 1080. You may change the video source by passing arguments while running the python script in console:
python face_detection_multithreading.py --source < video_source > --width < width_of_video_frames > --height < height_of_video_frames > -- model < path_to_facenet_model_pb_file >
Current script is made to run on CPU, to run on GPU comment line-2: os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
Current script will resize frame width to 400 before passing to Face Detection session, if you want to change that then you have to make changes to the method detect_faces in align module