ArcFace is an open source SOTA model for facial recognition. It has shown outstanding performance of 99.82% accuracy in LFW dataset. A detailed paper explanation is available in my article here.
To train arcface on any custom dataset:
- firstly the collected face dataset must be arranged in a particular format. Use LFW as reference.
- A text and csv file of name pairs would be generated using the script provided inside 'preparing_data' folder.
- Place the dataset in the path 'training/data/'
- change path to data and run the code in the notebook 'run.ipynb'
In this repo, The arcface model was trained on a small custom south asian face dataset and it yielded an accuracy of 98%