Neighborhood-Aware Attention Network for Semi-supervised Face Recognition

This repository is the official implementation of our IJCNN'20 paper Neighborhood-Aware Attention Network for Semi-supervised Face Recognition.

Requirements

• Python >= 3.5
• PyTorch >= 1.0.0
• torchvision >= 0.2.1
• numpy
• tqdm

To install requirements:

pip install -r requirements.txt

Data Preparation

• Download the full MS-Celeb-1M realeased by ArcFace from baidu or dropbox.
• Download the splitted image list produced by learn-to-cluster from GoogleDrive or OneDrive.
• Download the extracted features and precomputed knn for split0 (labeled) from GoogleDrive and move them to data/labeled.
• Download the extracted features and precomputed knn for split1 (unlabeled) from GoogleDrive and move them to data/unlabeled. We also provide the pretrained face recognition model (ArcFace trained using only split0) from GoogleDrive. You can use it to extract/compute features and knn files for your own data and move them to data/unlabeled. The structure of data is the same as:

  data
   ├── labeled
      ├── split0_feats.npz
      ├── split0_knn.npy
   ├── unlabeled
      ├── split1_feats.npz
      ├── split1_knn.npy
      ├── split1_labels.txt
  

Training

sh train_gat.sh

Please note that:

• We only use part0_train.list to train the GAT and classifier.
• You can download the pre-trained model weights from GoogleDrive．

Evaluation

1. Generate pseudo labels for unlabeled data

   sh eval.sh
   

   Please note that:

   • You should change the param model_path with your own or use the default setting.
   • After preparing your own data, you can change the params knn_path and feat_path in eval_gat.py to generate pseudo labels.

2. Evaluate with the true labels

   python statistics.py \
   pseudo_dir pseudl_data \
   split split1
   

   Please note that:

   • If your dirname of the generated pseudo label file is different, please replace pseudl_data with your own.
   • If your test set is not split1, please change split to your own data split.
   • It will output results, including BCubed Precision, BCubed Recall and BCubed F-measure.
   • The new pseudo_label_clean file will be saved into pseudo_dir by removing singleton clusters.

Joint Training Framework

You may refer to the following repository:
https://github.com/labyrinth7x/multi-task-face-recognition-framework

Results on split1_test

Method	Precision	Recall	F-score
NAAN	97.0	86.8	91.6

After removing singleton clusters

Method	Precision	Recall	F-score	Discard ratio (%)
NAAN	96.9	93.6	95.2	4.2

Citation

Please cite our paper if it helps your research:

@inproceedings{DBLP:conf/ijcnn/ZhangLL19,
 author    = {Qi Zhang, Zhen Lei, Stan Z.Li},
 title     = {Neighborhood-Aware Attention Network for Semi-supervised Face Recognition},
 booktitle = {IJCNN},
 year      = {2020}

Ackownledgement

The codes for pseudo label propagation are from CDP. Thanks for their work.