ChildPredictor

This is the official webpage of the paper "ChildPredictor: A Child Face Prediction Framework with Disentangled Learning", accepted to IEEE TMM, 2022

Arxiv: https://arxiv.org/abs/2204.09962

IEEE Xplore: https://ieeexplore.ieee.org/document/9749880

🚀 🚀 🚀 News:

May. 19, 2022: We release the data of the FF-Database, please see section 1.1 for the terms of use.
Apr. 15, 2022: We release the pre-trained models with validation samples for ChildPredictor.
Mar. 31, 2022: The paper is accepted by the IEEE Transactions on Multimedia.
Feb. 8, 2022: We release the code for ChildPredictor. We are considerring to release the original data of the collected FF-Database.

1 FF-Database

1.1 Data Download

If you would like to download the FF-Database data, please fill out an agreement to the FF-Database Terms of Use and send it to us at yzzhao2-c@my.cityu.edu.hk.

Please use the institutional email instead of anonymous addresses such as Gmail, QQMail, and hotmail. Make sure your email address is the same as it in the FF-Database Terms of Use.

1.2 Overview

The data collection pipeline is shown as follows:

Some families are shown as follows:

2 Results on Real Families

The generated results on the collected FF-Database:

The generated results on other datasets:

The disentangled learning analysis is as:

The ablation study is as:

3 Implementation

3.1 File Structure

Some files are not included in the current implementation since they are too large. The network architectures can be found in the code folder.

code
│
└───baby_model_pool
│   └───attgan
│   │   │   attgan_without_claloss_baby.pth
│   │   │   attgan_without_ganloss_celeba_baby.pth
│   │   │   attgan_without_ganloss_claloss_celeba_baby.pth
│   │   │   ...
│   └───inverse
│   │   │   Inverse_ProGAN_GAN_ACGAN_start-with-code.pth
│   │   │   Inverse_ProGAN_GAN_MSGAN_ACGAN_start-with-code.pth
│   │   │   Inverse_ProGAN_GAN_MSGAN_ACGAN_start-with-image.pth
│   │   │   ...
│   └───mapping
│   │   └───Mapping_Xencoder_full_ProGAN_GAN_MSGAN_ACGAN_deepArch_multi-gt_v4
│   │   │   │   MappingNet_Batchsize_32_Epoch_298.pth
│   │   └───Mapping_Xencoder_full_ProGAN_GAN_deepArch_multi-gt_v4
│   │   │   │   MappingNet_Batchsize_32_Epoch_298.pth
│   │   └───Mapping_Xencoder_wo-class_ProGAN_GAN_MSGAN_deepArch_multi-gt_v4
│   │   │   │   MappingNet_Batchsize_32_Epoch_298.pth
│   │   │   ...
│   └───ProGAN-ckp
│   │   │   ProGAN_pt_mixtureData_GAN.pth
│   │   │   ProGAN_pt_mixtureData_GAN_ACGAN.pth
│   │   │   ProGAN_pt_mixtureData_GAN_MSGAN.pth
│   │   │   ProGAN_pt_mixtureData_GAN_MSGAN_ACGAN.pth
│   │   │   ...
│
└───babyinverse (Ey)
│   │   ...
|
└───babymapping_1219 (T)
│   │   ...
│
└───Datasets
│   │   ...
│
└───ProGAN (Gy)
│   │   ...
│
└───AttGAN (please refer to AttGAN official webpage)
│   │   ...
│

3.2 Required Libraries

The implementation is based on CUDA 9.0 and PyTorch 1.1.0. The following packages are needed to be installed:

pytorch==1.1.0
torchvision==0.3.0
tensorflow-gpumkdir ./babymapping_1219/Models/pretrain
mv ./baby_model_pool/ProGAN-ckp/* ./babymapping_1219/Models/pretrain/
tensorboardx
pyyaml
tqdm
easydict

3.3 Testing a Real Face

First, download the pre-trained models from this link. It should be a large zip file with size of approximately 3.9 Gb.

After you have already downloaded the pre-trained models, enter code folder and unzip all the models under the ./code/baby_model_pool folder:

cd code
mkdir baby_model_pool
cd baby_model_pool
unzip Onedrive_baby_model_pool.zip
cd ..

Then, you need to move all the ProGAN pre-trained models under another path:

mkdir ./babymapping_1219/Models/pretrain
mv ./baby_model_pool/ProGAN-ckp/* ./babymapping_1219/Models/pretrain/

Next, you can test some validation samples (we have already put some examples under the code/babymapping_1219 folder):

cd babymapping_1219
python main.py

If you want to change the input images, see lines 38-39 of validation.yaml: https://github.com/zhaoyuzhi/ChildPredictor/blob/main/code/babymapping_1219/yaml/yaml/validation.yaml#L38-L39

3.4 Training

Currently, we do not release the full codes for training due to privacy issue.

3.5 Build Your Own Dataset

Please refer to code_FFDatabase_collection.

4 Network Architectures

5 Some Related Works

Zaman, Ishtiak and Crandall, David. Genetic-GAN: Synthesizing Images Between Two Domains by Genetic Crossover. European Conference on Computer Vision Workshops, 312--326, 2020.
Gao, Pengyu and Robinson, Joseph and Zhu, Jiaxuan and Xia, Chao and Shao, MIng and Xia, Siyu. DNA-Net: Age and Gender Aware Kin Face Synthesizer. IEEE International Conference on Multimedia and Expo (ICME), 2021.
Robinson, Joseph Peter and Khan, Zaid and Yin, Yu and Shao, Ming and Fu, Yun. Families in wild multimedia (FIW MM): A multimodal database for recognizing kinship. IEEE Transactions on Multimedia, 2021.

6 Citation

If you find this work useful for your research, please cite:

@article{zhao2022childpredictor,
  title={ChildPredictor: A Child Face Prediction Framework with Disentangled Learning},
  author={Zhao, Yuzhi and Po, Lai-Man and Wang, Xuehui and Yan, Qiong and Shen, Wei and Zhang, Yujia and Liu, Wei and Wong Chun-Kit and Pang, Chiu-Sing and Ou, Weifeng and Yu, Wing-Yin and Liu, Buhua},
  journal={IEEE Transactions on Multimedia},
  year={2022}
}

Please contact yzzhao2-c@my.cityu.edu.hk for further questions.

zhaoyuzhi / ChildPredictor