Keras-MORPH2-age-estimation

Keras implementation for MORPH2 dataset age estimation.

This project contains Mobilenet and Densenet with regression and DEX framework.

Update (2017/12/1)

• Fix inconsistent label problem.
• Add face align in preprocessing.

Update(2017/11/27)

• Change training default epoch to 90
• Decay learning rate at epoch [30,60]

How to run?

• Step.1 Download landmarks from https://github.com/xyfeng/average_portrait Unzip it and move to './landmarks'

• Step.2 Download MORPH2 dataset https://www.faceaginggroup.com/morph/ Unzip it under './morph2'

You have to apply for the dataset. No easy way to download it unfortunately :(

• Step.3 Preprocess the dataset (change isPlot inside TYY_MORPH_create_db.py to True if you want to see the process)

python TYY_MORPH_create_db.py --output morph_db.npz

• Step.4 Run the training and evalutation (change netType inside TYY_train_MORPH.py for different networks)

KERAS_BACKEND=tensorflow python TYY_train_MORPH.py --input ./morph_db.npz

Training and evaluation

• Training ratio: 0.8

• Validation ratio: 0.2

• Evaluation metric: Mean-absoluate-error (MAE) -> name: val_pred_a_mean_absolute_error

• Output example:

pred_a_softmax_loss: 2.4073 - pred_a_loss: 9.4221 - pred_a_softmax_acc: 0.1183 - pred_a_mean_absolute_error: 9.4221 - val_loss: 2.4423 - val_pred_a_softmax_loss: 2.4423 - val_pred_a_loss: 9.4864 - val_pred_a_softmax_acc: 0.1339 - val_pred_a_mean_absolute_error: 9.4864

Parameters

• DEX: num_neu is the output dimension of the classfication training part. Range of num_neu: [1~101]
• Mobilenet: alpha is the paramters to control the network size. Recommended value of alpha: 1, 0.5, 0.25
• Densenet: densenet_depth is the depth of the network (Obviously~~)

Dependencies

• Keras
• Tensorflow
• anaconda
• python3
• opencv3
• dlib
• moviepy
• pytables

References

• https://github.com/yu4u/age-gender-estimation
• https://github.com/titu1994/DenseNet
• R. Rothe, R. Timofte, and L. V. Gool, "Deep expectation of real and apparent age from a single image without facial landmarks," IJCV, 2016.
• https://github.com/fchollet/keras/blob/master/keras/applications/mobilenet.py
• https://github.com/xyfeng/average_portrait