This is the official implementation of "Weakly Supervised Face Naming with Symmetry-Enhanced Contrastive Loss" at WACV 2023
The packages used in our experiments are listed in requirement.txt
conda create -n facenaming python=3.7
source activate facenaming
pip install -r requirements.txt
We use the augmented LFW dataset as in Cross-Media Alignment of Names and Faces
If you cannot find the dataset online, please contact us. The dictionary for reading the dataset is provided in /Berg
You can find Celeberty Together here.
You can download the dictionary I made for Celebrity Together here.
mkdir CelebrityTo # make the dir and download the dict there
You can train SECLA model using the command
bash run_face_align.sh # LFW data
bash run_face_align_celeb.sh # Celebrity Together data
Remember to replace OUTPUTDIR and DATADIR accordingly.
You can train SECLA-B model using the command
bash run_face_align_incre.sh # LFW data
bash run_face_align_celeb_incre.sh # Celebrity Together data
Remember to replace OUTPUTDIR and DATADIR accordingly.
If you are interested in using CharacterBERT as backbone, please check here. You can clone it under /models directory.
You can calculate metrics using
bash run_test.sh
I'll further clean the code when I have time. Thanks for your understanding:)
If you are interested in our work, please cite it as:
@InProceedings{Qu_2023_WACV,
author = {Qu, Tingyu and Tuytelaars, Tinne and Moens, Marie-Francine},
title = {Weakly Supervised Face Naming With Symmetry-Enhanced Contrastive Loss},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
month = {January},
year = {2023},
pages = {3505-3514}
}