AgeGenderPrediction

These models are based on this paper

Network Architecture

The network contains three convolutional layers, each followed by a rectified linear operation and pooling layer. The first two layers also follow normalization using local response normalization. The first Convolutional Layer contains 96 filters of 7×7 pixels, the second Convolutional Layer contains 256 filters of 5×5 pixels, The third and final Convolutional Layer contains 384 filters of 3×3 pixels. Finally, two fully-connected layers are added, each containing 512 neurons.

Dataset Preparation

Download

The dataset used for training is UTKFace but filter to get only Asian people for the project’s purpose. Full UTKFace dataset can be download at UTKFace

To download UTKFace AsianOnly:

./download.sh

Split train/test and oversampling train set

python3 build_dataset.py

Note: the split ratio is 0.8. After spliting into train/test set, the augmentation process is then implemented by flipping and adding noise to original images to balance number of sample between age classes.

Number of image per age class:
0-9    10-19    20-29    30-39    40-49    50-59    60-69    70-79    80-120
--------------------------------------------------------------
734    78    1066    423    126    95    76    67    82
Number of image per gender class:
Male    Female
--------------
1260    1487
AFTER OVERSAMPLING:
Number of image per age class:
0-9    10-19    20-29    30-39    40-49    50-59    60-69    70-79    80-120
--------------------------------------------------------------
1066    1066    1066    1065    1066    1065    1066    1065    1066
Number of image per gender class:
Male    Female
--------------
5690    3901

Create lmdb - Make mean file

./create_lmdb.sh
./create_mean.sh

Train and Test

Train

Train AgeNet

./train_AgeNet.sh

Train GenderNet

./train_GenderNet.sh

Test

Test AgeNet

./test_AgeNet.sh

Test GenderNet

./test_GenderNet.sh

Test result

AgeNet

Accuracy: 66.5%, 457/687 corrects
Confusion matrix:
Actual\Predict    0-9    10-19    20-29    30-39    40-49    50-59    60-69    70-79    80-120    
        -----------------------------------------------
0-9        178    2    7    2    0    0    0    0    2
10-19        4    2    9    3    1    0    0    0    0
20-29        1    2    214    44    2    2    2    2    0
30-39        2    2    50    37    8    5    1    2    1
40-49        0    1    6    10    8    2    1    0    0
50-59        0    0    1    4    3    3    5    1    0
60-69        2    0    3    2    0    3    3    3    0
70-79        0    0    1    0    0    4    8    4    3
80-120        1    0    0    2    0    0    2    6    8

GenderNet

Accuracy: 85.2%, 585/687 corrects
Confusion matrix:
Actual\Predict    Male    Female    
        -----------------------------------------------
Male        267    53
Female        49    318

Convert to NCS graph

This script compiles AgeNet and GenderNet caffe models into NCS graphs. It also converts mean file.

./compile2Movidius.sh

Run inference

python3 run_inference.py --mean_file age_gender_mean.npy --image_path image/test_0.png

Citation

@inproceedings{LH:CVPRw15:age,
 author    = {Gil Levi and Tal Hassner},
 title     = {Age and Gender Classification Using Convolutional Neural Networks},
 booktitle = {IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) workshops},
 month    =  {June},
 year     = {2015},
 URL     = {\url{https://osnathassner.github.io/talhassner/projects/cnn_agegender}}
}