README.md

Introduction

This is the official PyTorch implementation of the NeurIPS 2023 paper.

UniTSFace: Unified Threshold Integrated Sample-to-Sample Loss for Face Recognition.pdf

Supplementary.pdf

Get started

Requirement: PyTorch >= 1.8.1

1. Prepare dataset

   Download CASIA-Webface preprocessed by insightface.

   unzip faces_webface_112x112.zip

2. Train model

   Modify the 'data_path' in train.py (Line 57)

   Select and uncomment the 'sample_to_sample_loss' in backbone.py (Line 71)

   python train.py

3. Test model

   python pytorch2onnx.py
   zip model.zip model.onnx

   Upload model.zip to MFR Ongoing and then wait for the results.

   We provide a pre-trained model (ResNet-50) on Google Drive for easy and direct development. This model is trained on CASIA-WebFace and achieved 50.25% on MR-All and 99.53% on LFW.

Citation

If you find UniTSFace useful in your research, please consider to cite:

@InProceedings{NeurIPS_2023_UniTSFace,
  author    = {Li, Qiufu and Jia, Xi and Zhou, Jiancan and Shen, Linlin and Duan, Jinming},
  title     = {UniTSFace: Unified Threshold Integrated Sample-to-Sample Loss for Face Recognition},
  journal   = {Advances in Neural Information Processing Systems},
  volume    = {36},
  pages     = {32732--32747},
  year      = {2023}
}