Face Embeddings

This repository integrates and extends the code from https://github.com/jonasgrebe/pt-femb-face-embeddings.

The datasets for training can be downloaded from the Insightface Datasets Github Page. The training datasets are provided in mxnet RecordIO file format.

Set up repository

1. Install miniconda

2. Create environment with

   conda env create -n femb -f environment.yml

3. Install pip requirements with

   conda activate femb
   pip install -r requirements.txt

4. Install pre-commit hooks with

   pre-commit install

Run training and inference

We provide a train script and a predict script that can be executed with a matching config file.

Example for training:

python train.py --config configs/train_arcface_ddp.yaml

Example for prediction:

python predict.py --config configs/predict_arcface.yaml

To generate predictions on the official model checkpoints, you first need to download the checkpoints and put them into the checkpoints/ directory.

Checkpoints for ArcFace can be downloaded at: Arcface Repository

Checkpoints for MagFace can be downloaded at: MagFace Repository

Model references

Backbone

• IResNet: Improved Residual Networks for Image and Video Recognition (ICCV, 2021)

Headers

• SphereFaceHeader: SphereFace: Deep Hypersphere Embedding for Face Recognition (CVPR, 2017)
• CosFaceHeader: CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR, 2018)
• ArcFaceHeader: ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR, 2019)
• MagFaceHeader: MagFace: A Universal Representation for Face Recognition and Quality Assessment (CVPR, 2021)

Todos

• Load ArcFace checkpoint from official Arcface repository
• Train own ArcFace model and save it compatible to official Arcface checkpoint
• Add ElasticFace header
• Compare MagFace training to official Magface code
• Remove mxnetdependency. Therefore it is necessary to convert the datasets.