CVPR2021 'Pseudo Facial Generation with Extreme Poses for Face Recognition'

Requirements
Experiments on MultiPIE dataset

Requirements

For finetuning LightCNN

Considering that the original LightCNN_V2is based on an old environment, so we need create one based on CUDA9.2.

cuda92
numpy
opencv
python=2.7
pytorch=0.4.1
scikit-learn
scipy
torchvision=0.2.1
tqdm

For training the generator

cudatoolkit=10.0
numpy
opencv
python=3.6
pytorch=1.4.0
scikit-learn
scipy
torchvision=0.5.0
tqdm

You can also try some other versions (Most of GPUs are now Nvidia 30s, and they do not support CUDA lower than 11.0), but you may encounter some dependency problems which can be easily fixed by changing some out-of-date functions.

Models

LightCNN

lightCNN_152_checkpoint.pth.tar

Datasets

MultiPIE

Experiments on MultiPIE dataset

Due to the copyright, we can only offer a script to test the pretrained models easily, also you can train models from scratch if you want. So for example assume you want to test it when the angle is 90 $\degree$ in our setting2, you should do the following:

In experiment-on-multipie/pre_train folder, you should put the LightCNN_V2 model which is named as lightCNN_152_checkpoint.pth.tar. In experiment-on-multipie/model folder, you should put the generator model which is named as netG_model_epoch_80_iter_0.pth.

Then, you should go into experiment-on-multipie/run_evaluation.sh and check the path of data and model. Depending on your device, you can adjust the batch size and GpuID as you like. --probe_list depend what angle you are testing. In this example, we should choose multipie_90_test_list.txt in the MultiPIE data folder.

Finally, you just run the script we offer:

sh run_evaluation.sh 80

And the results should be like this:

For any other angels, you can test them easily by changing the probe_list file. For setting1, you can change the data file to.

Citation

If you find this project useful in your research, please consider citing:

@inproceedings{DBLP:conf/cvpr/WangM0L021, author = {Guoli Wang and Jiaqi Ma and Qian Zhang and Jiwen Lu and Jie Zhou}, title = {Pseudo Facial Generation With Extreme Poses for Face Recognition}, booktitle = {{IEEE} Conference on Computer Vision and Pattern Recognition, {CVPR} 2021, virtual, June 19-25, 2021}, pages = {1994--2003}, year = {2021}, }

Future Work

More exciting researches are under constructions. We will release them later.

Contacts

If you have any questions, you can send email to wangguoli1990@mail.tsinghua.edu.cn and jiaqima@whu.edu.cn.

leonmakise/PFGEPFR