SECLA & SECLA-B (WACV 2023)

This is the official implementation of "Weakly Supervised Face Naming with Symmetry-Enhanced Contrastive Loss" at WACV 2023

Preperation

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

Data

Augmented LFW

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

Celebrity Together

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

Train

SECLA

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.

SECLA-B

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.

CharacterBERT

If you are interested in using CharacterBERT as backbone, please check here. You can clone it under /models directory.

Test

You can calculate metrics using

bash run_test.sh

Note

I'll further clean the code when I have time. Thanks for your understanding:)

Citation

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}
}