High Speed Real-Time Face-Age-Gender-Recognition

• A real-time age and gender recognition of face model with only 1mb, 0.531×10^6 parameters and 0.074 GFLOPs.
• Using OpenVINO to accelerate the speed of real-time recognition on intel D415.
• The backbone of model structure is MobileFaceNet.
• CelebA, AFAD, MegaAge-Asian and UTKFace dataset is used for model training.

Code Author: Vaan Lin

Last update: 2020/07/22

Face image inference demo

​ ​ ​

Requirements

Please install Anaconda(Python 3.7 64-Bit) first.

NVIDIA

• CUDA 10.1
• cudnn 7.6.4

Python

• Tensorflow-GPU 2.2.0
• Keras 2.3.1
• cvlib 0.2.2
• numpy 1.17.4
• pyrealsense2 2.31.0.1235
• matplotlib 3.3.3
• opencv-python 4.5.1.48

You can install the package of above at once by using command:

pip install -r requirements.txt

OpenVINO

• OpenVINO the lastest version

Dependencies

• Anaconda
• MobileFaceNet (https://arxiv.org/ftp/arxiv/papers/1804/1804.07573.pdf)
• cvlib (https://github.com/arunponnusamy/cvlib)

Pipeline

• This part describes the roughly process of TFrecords generation, training and model convertion.

1. Dataset --> ./TFRecords_Create/gen_TFRecords.py --> [Asian and UTK].tfrecords
2. [Asian and UTK].tfrecords, MFN_62_075_gender_pre-trained.h5 --> ./Training/MFN_Train.py
3. ./Training/Results/Keras_h5/MFN.h5 --> ./Tools_Convert/Keras2pb.py --> MFN.pb
4. MFN.pb --> Convert pb to IR --> MFN.bin, MFN.mapping and MFN.xml

Datasets

If you wanna use your datasets to train on my model, please notice the following:

1. This model is for age and gender recognition, so make sure your dataset can be labeled like:

​

2. When you have already prepared your label.csv, put your images at

./FaceAgeGenderRecognition/Datasets

and you can generate TFRecors by executing gen_TFRecords.py

3. The results of TFRecords will save at

./FaceAgeGenderRecognition/TFRecords_Create/TFRecords/train(or test)

Simple Steps of Execution

• If you wanna use the model to inference your images

1. Put your images in

./FaceAgeGenderRecognition/Demo/Image/Demo_Image

2. Double click run_Image_Demo.bat
3. The results will be saved in

./FaceAgeGenderRecognition/Demo/Image/Results

Insure you have already installed OpenVINO.

• If you wanna use the model to achieve the real-time recognition on intel D415

1. Double click run_Webcam_Demo.bat

• If you want to re-train the model of using your dataset of TFRecords

1. Double click run_MFN_Train.bat

Command of Code Execution

Insure you have already installed OpenVINO.

• If you wanna use the model to inference your images

1. Activate your Anaconda environment
2. Put your images in

./FaceAgeGenderRecognition/Demo/Image/Demo_Image

3. Executed by python:

[OpenVINO]
python ./Demo/Image/Image_Test.py --m OpenVINO --i ./Demo/Image/Demo_Image --o ./Demo/Image/Results/ --x ./Training/Results/Openvino_IR/MFN.xml --b ./Training/Results/Openvino_IR/MFN.bin

[Keras]
python ./Demo/Image/Image_Test.py --m Keras --i ./Demo/Image/Demo_Image --o ./Demo/Image/Results/ --h [Path of .h5 file]

--m means what model type you wanna use
--x means path of OpenVINO .xml file
--b means path of OpenVINO .bin file
--i means path of input images
--o means path of output images

• If you wanna use the model to achieve the real-time recognition on intel D415

1. Activate your Anaconda environment
2. Executed by python:

python ./Demo/RealTime/Webcam.py --x ./Training/Results/Openvino_IR/MFN.xml --b ./Training/Results/Openvino_IR/MFN.bin --r 1280

--x means path of OpenVINO .xml file
--b means path of OpenVINO .bin file
--r means the resolution of D415(include 1920(1920x1080), 1280(1280x720) and 960(960x540))

• If you want to create the tfrecords by yourself

1. Activate your Anaconda environment
2. Executed by python(Asian):

python ./TFRecords_Create/gen_TFRecords.py --i ./FaceAgeGenderRecognition/Datasets/Asian_FaceData/ --c ./TFRecords_Create/Asian_FaceData.csv --t ./TFRecords_Create/TFRecords --n Asian

--i means the path of dataset
--c means the path of csv file
--t means the path of output tfrecords
--n means the name of output tfrecords

• If you want to re-train the model of using your dataset of TFRecords

1. Activate your Anaconda environment
2. cd ./FaceAgeGenderRecognition/Training
3. Executed by python:

python MFN_Train.py --tn ./FaceAgeGenderRecognition/Outputs/TFRecords/test/ --m .././Model/Backbone/MFN_62_075_gender_pre-trained.h5 --p Y

--tn means path of training tfrecords
--ts means path of testing tfrecords
--m  means the path of model
--p  means If use pre-trained model or not

Training Details

All of training details you can check the document named FaceAgeGenderRecognition.pdf.

Results

Compared with OpenVINO and InsightFace

• Gender

​ ​

• Age

Results of (a) OpenVINO, (b) InsightFace and (c) Our Model age estimation. X-axis means range of age from 18 to 65, and y-axis means the mean absolute error (MAE) of model estimation in every age.

​ ​ ​