Update 10/31/2019 We upload the loss.py as our implementation of the original ArcFace and the proposed ArcNegFace. Now you can use them to train the model and compare the results. If you use the ms1m-v3 provided by the LFR challenge, you would reproduce the reported results as listed in our paper.
This is the model and code for the paper 'Towards Flops-constrained Face Recognition', which win the 1st place in the ICCV19 Lightweight Face Recognition Challenge, large video track. This repo only covers the two 30Gflops tracks (deepglint-large and iQiyi-large) that we took participate in. For more details such as network design, training strategy and bag of tricks, please refer to our workshop paper.
- Using this link to download the EfficientPolyFace and QAN++ models.
- Clone this repo and override the files with the same name.
- Use this example to generate the representation (feature vector) for a single image:
python test_tp.py --test_img 00000f9f87210c8eb9f5fb488b1171d7.jpg --save_path ./
- Use this example to generate the representation and the quality for a single image. This will produce two files, which store the representation and quality respectively:
python test_quality.py --test_img 000005_99.jpg
Note that to generate quality for a set/sequence/video, you need to generate representations and qualities for all frames in this step by a loop.
- Use
generate_aggregation_feat.pyto generate presentations for a video.
Please cite our paper if this project helps your research.
@inproceedings{liu2019towards,
title={Towards Flops-constrained Face Recognition},
author={Liu, Yu and Song, Guanglu and Zhang, Manyuan and Liu, Jihao and Zhou, Yucong and Yan, Junjie},
booktitle={Proceedings of the ICCV Workshop},
year={2019}
}