Confidence-Calibrated Face and Kinship Verification

💡 This is the official implementation of the paper "Confidence-Calibrated Face and Kinship Verification" [arxiv]

This work has been accepted by IEEE Transactions on Information Forensics and Security (TIFS).

We introduce a simple yet effective confidence measure for face and kinship verification, which allows the verification models to transform the similarity score into a confidence score for a given face pair. We propose a confidence-calibrated approach called Angular Scaling Calibration (ASC). ASC is easy to implement and can be directly applied to existing face and kinship verification models without model modifications, yielding accuracy-preserving and confidence-calibrated probabilistic verification models. In addition, we model the uncertainty of calibrated confidence to enhance the robustness of the verification models to noisy data.

📋 This Repository Can Do

1. Confidence Measure ✔️
2. Confidence Calibration ✔️
3. Uncertainty Estimation ✔️

🔧 Installation

To run experiments, first clone the repository and install requirements.txt.

$ git clone https://github.com/cnulab/ASC.git
$ cd ASC
$ pip install -r requirements.txt

You can download the FIW dataset and pre-trained models via "Baidu Cloud" (pwd 6789) or "Google Drive".
Copy the model files to the models and the FIW dataset to the data.
You can get the following directory structure.

    |--data                         
        |--FIW           
            |--pairs
            |--Test
            |--Train
            |--Validation           
    |--models
        |--resnet50.pth
        |--resnet101.pth

🚀 Finetune

You can finetune a pre-trained ResNet101 network using ArcFace loss:

$ python train_backbone.py --loss arcface --backbone resnet101  --optimizer adam  --batch_size 32  --s 10.0 --m 0.4 --image_size 112

The checkpoint will be saved in the arcface_resnet101. More commands (other losses and models) in run.sh.

🔑 Calibration

Calibration using ASC:

$ python calibration.py  --backbone resnet101  --model_root arcface_resnet101 --image_size 112

Use the --model_root to set the path of your checkpoint.

🌐 Uncertainty Estimation

Learn an uncertainty prediction head using the following command:

$ python train_uncertainty.py  --backbone resnet101  --model_root arcface_resnet101 --image_size 112

The checkpoints will be saved in the arcface_resnet101/heads. Then use the following command for uncertainty estimation:

$ python eval_uncertainty.py  --backbone resnet101  --model_root arcface_resnet101 --image_size 112 --head_checkpoint_path 'heads/epoch_60_head.pkl'

🔗 Citation

If this work is helpful to you, please cite it as:

@ARTICLE{xu2023confidence,
  author={Xu, Min and Zhang, Ximiao and Zhou, Xiuzhuang},
  journal={IEEE Transactions on Information Forensics and Security}, 
  title={Confidence-Calibrated Face and Kinship Verification}, 
  year={2023},
  doi={10.1109/TIFS.2023.3318957}
}