Age and gender estimation based on Convolutional Neural Network and TensorFlow

---

UPDATE: There are some compatibility issues under python3 reported by others, recommend using python2 or you should adjust some codes manually

This is a TensorFlow implement of face age and gender estimation which first using dlib to detect and align faces in the picture and then using a deep CNN to estimate age and gender.As you can see below,this project can estimate more than one face in a picture at one time.

Dependencies

This project has following dependencies and tested under CentOS7 with Python2.7.14

• tensorflow==1.4
• dlib==19.7.99
• cv2
• matplotlib==2.1.0
• imutils==0.4.3
• numpy==1.13.3
• pandas==0.20.3

Usage

Make tfrecords

In order to train your own models,you should first download imdb or wiki dataset,and then extract it under data path,after that,images path should look like

/path/to/project/data/imdb_crop/00/somepictures
/path/to/project/data/imdb_crop/01/somepictures
....
/path/to/project/data/imdb_crop/99/somepictures

Then you can run

python convert_to_records_multiCPU.py --imdb --nworks 8

to convert images to tfrecords.--imdb means using imdb dataset,--nworks 8 means using 8 cpu cores to convert the dataset parallelly.Because we will first detect and align faces in the pictures,which is a time consuming step,so we recommend to use as many cores as possible.Intel E5-2667 v4 and with 32 cores need approximately 50 minutes.

Train model

Once you have converted images to tfrecords,you should have the following path:

/path/to/project/data/train/train-000.tfrecords
......
/path/to/project/data/test/test-000.tfrecords
......

At present,our deep CNN uses FaceNet architecture,which based on inception-resnet-v1 to extract features.To speed up training,we use the pretrained model's weight from this project and have converted the weight to adapt our model,you can download this converted pretrained facenet weight checkpoint from here or here.Extract it to path models.

/path/to/project/models/checkpoint
/path/to/project/models/model.ckpt-0.data-00000-of-00001
/path/to/project/models/model.ckpt-0.index
/path/to/project/models/model.ckpt-0.meta

NOTE: This step is optional,you can also train your model from scratch. To start training,run

python train.py --lr 1e-3 --weight_decay 1e-5 --epoch 6 --batch_size 128 --keep_prob 0.8 --cuda

NOTE: Using the flag --cuda will train the model with GPU.

Using tensorboard to visualize learning

tensorboard --logdir=./train_log

Test model

You can test all your trained models on testset through

python test.py --images "./data/test" --model_path "./models" --batch_size 128 --choose_best --cuda

Flag --cuda means using GPU when testing.--choose_best means testing all trained models and return the best one.If you just want to test the latest saved model,without this flag.

python test.py --images "./data/test" --model_path "./models" --batch_size 128 --cuda

One picture demo

If you just want to test the model on your own picture,run

python eval.py --I "./demo/demo.jpg" --M "./models/" --font_scale 1 --thickness 1

Flag --I tells where your picture is.If the text label too small or too large on the picture,you can use a different --font_scale 1 and --thickness 1 to adjust the text size and thickness. We also provide a pretrained model,you can download from here or here and extract it to models path.

Picture from web cam

First download the pretrained model from here or here and extract to models path. In order to get pictures from web cam, you may need to uninstall your cv2 and install it from source if have problems when running the below command:

python demo.py 

TODO

• First version of project
• Code review and add some commits
• Add a readme doc
• Add a demo using pictures from web cam

References and Acknowledgments

This project is a part of my class project of Machine Learning course at Zhejiang University,following papers and codes are referred:

1. Rothe R, Timofte R, Van Gool L. Dex: Deep expectation of apparent age from a single image[C]//Proceedings of the IEEE International Conference on Computer Vision Workshops. 2015: 10-15.
2. Rothe R, Timofte R, Van Gool L. Deep expectation of real and apparent age from a single image without facial landmarks[J]. International Journal of Computer Vision, 2016: 1-14.
3. IMDB-WIKI – 500k+ face images with age and gender labels
4. yu4u/age-gender-estimation
5. davidsandberg/facenet
6. Face Alignment with OpenCV and Python