UnsupervisedFaceClustering

Evaluation Instruction

    pip install -r requirements.txt

    In eval.sh uncomment which version to use.
    3 flags:
        - model_version: which model version to use.
        - dataset_name: this is the name of dataset, default custom. 
        - dataset_folder: path of the folder where all images are there.
        (Note: dataset_folder here is the folder where images are stored inside given classes name)
    output will be stored here: "./evaluation_results/{model_version}/{dataset_name}/"

    If want to test on 105 and LFW, download the data from link and put it inside ./Data/ Folder.

Downloads

• 105_test
• lfw
• evaluation_results

Using Docker:

    docker build -t face_clust .
    docker run -v ./evaluation_results/:/app/evaluation_results/ face_clust

Dataset

Dataset Index	Dataset	Number of Identities	Total Number of Images
1.	105_classes_pins_dataset_train	59	9,848
2.	105_classes_pins_dataset_test	46	7,686
3.	CASIA	10,575	452,627
4.	CELEBA	10,177	201,172
5.	DigiFace	10,000	720000
6.	LFW	311	5425

Model Training

Backbone	Precision	Training Dataset	Evaluation Metrics	Model Path	Model Size	Model FPS
MobileFaceNet	FP32	105_train + CELEBA + DigiFace + CASIA	Dataset NMI ARI Pairwise F1 105_test 0.907 0.856 0.862 lfw 0.965 0.955 0.956		4.2 mb	124
MobileFaceNet	FP16	105_train + CELEBA + DigiFace + CASIA	Dataset NMI ARI Pairwise F1 105_test 0.917 0.867 0.873 lfw 0.965 0.955 0.956		2.2 mb	1000
MobileFaceNet	PTQ(INT8)	105_train + CELEBA + DigiFace + CASIA	Dataset NMI ARI Pairwise F1 105_test 0.910 0.860 0.865 lfw 0.967 0.955 0.956		1.6 mb	1300
MobileFaceNetv2	FP32	105_train + CELEBA + DigiFace + CASIA	Dataset NMI ARI Pairwise F1 105_test 0.876 0.809 0.815 lfw 0.961 0.949 0.951		4.8 mb	105
MobileFaceNetv2	FP16	105_train + CELEBA + DigiFace + CASIA	Dataset NMI ARI Pairwise F1 105_test 0.879 0.817 0.822 lfw 0.96 0.948 0.95		2.6 mb	938
MobileFaceNetv2	PTQ(INT8)	105_train + CELEBA + DigiFace + CASIA	Dataset NMI ARI Pairwise F1 105_test 0.802 0.869 0.795 lfw 0.96 0.946 0.947		2 mb	1133

Sample Results (Test Data)

Model	105_Classes(Kaggle)	LFW
MobileFaceNet
MobileFaceNetv2

Umap Visualization (Test Data)

Model	105_Classes(Kaggle)	LFW
MobileFaceNet
MobileFaceNetv2