Age and Gender Estimation

This is a Keras implementation of a CNN network for estimating age and gender from a face image. In training, the IMDB-WIKI dataset is used.

Dependencies

Tested on Ubuntu 16.04, Python 3.5.2, CUDA 8.0, cuDNN 5.0.

• Python3.5+
• Keras
• scipy, numpy, Pandas, tqdm
• OpenCV3

Usage

Download the dataset

The dataset is downloaded and extracted to the data directory.

./download.sh

Create data

Filter out noise data and serialize images and labels for training into .mat file. Please check check_dataset.ipynb for the details of the dataset.

python3 create_db.py --output data/imdb_db.mat --db imdb --img_size 64

Train network

Train the network using the training data created above.

python3 train.py --input data/imdb_db.mat

Plot training curves from history file.

python3 plot_history.py -input models/history_16_8.h5 

Network architecture

In the original paper, the pretrained VGG network is adopted. Here the Wide Residual Network (WideResNet) is trained from scratch. I modified the @asmith26's implementation of the WideResNet; two classification layers (for age and gender estimation) are added on the top of the WideResNet. Note that age and gender are estimated independently using different two CNNs.

Results

Trained on imdb, tested on wiki.