face-clustering

This face clustering makes use of existing work([CDP codes]https://github.com/XiaohangZhan/cdp). This clustering algorithm has been submitted to...

Paper

1. Consensus-Driven Propagation in Massive Unlabeled Data for Face Recognition, ECCV 2018 [Project Page]
2. Linkage-based Face Clustering via Graph Convolution Network, CVPR 2019
3. Learning to Cluster Faces on an Affinity Graph, CVPR 2019 [Project Page]
4. Efficient Parameter-free Clustering Using First Neighbor Relations, CVPR 2019

Requirements

• Python >= 3.6
• scikit-learn

Setup and get data

Install dependencies

pip install -U scikit-learn

Please download and use this dataset DATASET.md to evaluate the clustering algorithm. You may also use the face embeddings dataset from (https://github.com/Zhongdao/gcn_clustering).

Run

git clone [email protected]:XiaohangZhan/cdp.git

Put your face embeddings file e.g. part0.bin in the folder.

cdp/data/unlabeled/omni/features/part0.bin

Create the config.yaml file. Cluster using the commmands in (https://github.com/XiaohangZhan/cdp).

python -u main.py --config experiments/omni/config.yaml

Download this github python files *.py into the folder

cdp/data/unlabeled/omni/

Create a list.txt file from the meta.txt file using our codes

python create_list.py

Then further split the clusters using our codes

python cluster_finetuning_and_perf_other_algo.py

Remember to set the data_name and features to the values below

#load features
data_name='unlabeled/omni'
feats1 = load_feats('C:/Users/isetsk/Downloads/cdp-master/cdp-master/data/{}/features/{}.bin'.format(data_name, 'part0'),256)

And set the knn file to

knn1=np.load('./knn/part0_k15.npz')

Put the ground truth labels meta.txt and cdp clustering labels meta.txt in the correct folder and set e.g.

label_true=meta=np.loadtxt('meta.txt')
label_predict=np.loadtxt('..../cdp/experiments/omni/output/k15_vote_accept0_th0.7/sz600_step0.05/meta.txt')

Then label Propagation of Remaining Unlabeled Face embeddings using

python label_pred_smooth_by_neighbors.py

Results

Look at the results in our paper (... links to our paper).

Citations

If you use our work, please cite

@inproceedings{ourpaper,
  title={,
  author={},
  booktitle={},
  year={}
}