FaceGCD: Generalized Face Discovery via Dynamic Prefix Generation [BMVC2025]

Official PyTorch implementation of "FaceGCD: Generalized Face Discovery via Dynamic Prefix Generation".

📄 Paper: arXiv:2507.22353

---

📋 Overview

Recognizing and differentiating among both familiar and unfamiliar faces is a critical capability for face recognition systems and a key step toward artificial general intelligence (AGI). This repository introduces Generalized Face Discovery (GFD), a novel open-world face recognition task that unifies traditional face identification with generalized category discovery (GCD).

Key Features

• Novel Task: GFD requires recognizing both labeled and unlabeled known identities (IDs) while simultaneously discovering new, previously unseen IDs
• Dynamic Prefix Generation: Instance-specific feature extractors using lightweight, layer-wise prefixes generated on-the-fly by a HyperNetwork
• State-of-the-art Performance: Significantly outperforms existing GCD methods and ArcFace baseline on fine-grained face recognition tasks
• High Cardinality Support: Handles hundreds or thousands of visually similar face IDs effectively

---

🎯 Key Contributions

1. Generalized Face Discovery (GFD): A new task formulation that bridges face identification and clustering in open-world scenarios
2. Dynamic Prefix Mechanism: HyperNetwork-based prefix generators that create instance-specific feature extractors without massive model capacity
3. Comprehensive Benchmarks: Six GFD benchmark datasets (YTF-500/1000/2000, CASIA-500/1000/2000)
4. Strong Generalization: Competitive performance on generic GCD benchmarks (CIFAR-100, ImageNet-100, CUB, etc.)

---

🚀 Installation

Requirements

• Python 3.8+
• PyTorch 2.3.1+
• CUDA 11.8+
• APEX (for mixed precision training)

Setup

1. Clone the repository:

git clone https://github.com/yourusername/FaceGCD.git
cd FaceGCD

2. Create a conda environment:

conda create -n facegcd python=3.8
conda activate facegcd

3. Install dependencies:

pip install -r requirements.txt

4. Install APEX (optional but recommended for faster training):

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" ./

---

📊 Dataset Preparation

GFD Benchmark Datasets

We provide six benchmark datasets for evaluating the Generalized Face Discovery task:

Dataset	Known IDs	Unknown IDs	Train Samples	Test Samples
YTF-500	250	250	48,089	11,779
YTF-1000	500	500	96,002	23,523
YTF-2000	1,000	1,000	190,248	46,615
CASIA-500	250	250	46,991	11,999
CASIA-1000	500	500	89,508	22,867
CASIA-2000	1,000	1,000	184,432	47,114

Download

Datasets and pretrained checkpoints will be available soon via Google Drive.

Once downloaded, organize the data as follows:

FaceGCD/
├── youtube_faces_500/
│   ├── train/
│   └── test/
├── youtube_faces_1000/
│   ├── train/
│   └── test/
├── youtube_faces_2000/
│   ├── train/
│   └── test/
├── casia_webface_500/
│   ├── train/
│   └── test/
├── casia_webface_1000/
│   ├── train/
│   └── test/
└── casia_webface_2000/
    ├── train/
    └── test/

Pretrained Weights

Download the DINO pretrained weights (checkpoint.pth) and place them in the project root directory.

---

🏋️ Training

Train on YouTube Faces 1000

bash shell/train_youtube1000.sh

Or run directly:

CUDA_VISIBLE_DEVICES=0,1 torchrun \
    --nproc_per_node=2 \
    --master_port=55411 \
    train.py \
    --pretrained \
    --return-embed \
    --save-images \
    --pin-mem \
    --layer-embed \
    --experiment gcd_youtubefaces_1000_part_fvit_norm_prefix10 \
    --amp \
    --prefix_tuning \
    --prefix_length 10 \
    --data-path youtube_faces_1000 \
    --dataset youtubefaces_1000 \
    --amp-impl apex \
    --pretrained_weights checkpoint.pth \
    --log-wandb \
    --patch_size 8 \
    --input-size 3 112 112

Key Training Arguments

• --prefix_tuning: Enable dynamic prefix generation
• --prefix_length: Number of prefix tokens (default: 10)
• --dataset: Choose from youtubefaces_500, youtubefaces_1000, youtubefaces_2000, casia_500, casia_1000, casia_2000
• --data-path: Path to dataset directory
• --pretrained_weights: Path to DINO pretrained checkpoint
• --log-wandb: Enable Weights & Biases logging
• --amp: Enable mixed precision training

Training on Other Datasets

For CASIA-WebFace or different scales:

# CASIA-1000
python train.py \
    --dataset casia_1000 \
    --data-path casia_webface_1000 \
    --prefix_length 10 \
    --prefix_tuning \
    [other arguments...]

# YTF-500
python train.py \
    --dataset youtubefaces_500 \
    --data-path youtube_faces_500 \
    --prefix_length 10 \
    --prefix_tuning \
    [other arguments...]

---

🔍 Feature Extraction

After training, extract features for evaluation:

bash shell/feature_extract_youtube1000.sh

Or run directly:

CUDA_VISIBLE_DEVICES=0,1 torchrun \
    --nproc_per_node=2 \
    --master_port=25411 \
    extract_features.py \
    --pretrained \
    --return-embed \
    --save-images \
    --pin-mem \
    --experiment gcd_youtubefaces_1000_part_fvit_pretrain_prefix10 \
    --amp \
    --prefix_tuning \
    --prefix_length 10 \
    --data-path youtube_faces_1000 \
    --dataset youtubefaces_1000 \
    --amp-impl apex \
    --save_dir results \
    --landmark_cnn \
    --pretrained_weights checkpoint.pth \
    --checkpoint_weights model_best.pth.tar \
    --patch_size 8 \
    --input-size 3 112 112

This will save extracted features to the results/ directory.

---

📈 Evaluation

Perform Semi-Supervised K-Means clustering on extracted features:

bash shell/semi_supervised_k_means_youtube1000.sh

Or run directly:

CUDA_VISIBLE_DEVICES=0,1 torchrun \
    --nproc_per_node=2 \
    --master_port=32411 \
    SSK.py \
    --experiment GCD_base_prefix_gen \
    --dataset youtubefaces_1000 \
    --K 1000 \
    --max_kmeans_iter 500 \
    --k_means_init 10 \
    --experiment_idx gcd_youtubefaces_1000_part_fvit_pretrain_prefix10 \
    --save_dir results \
    --data-path youtube_faces_1000

Evaluation Metrics

• ACC (Clustering Accuracy): Overall clustering accuracy

---

📊 Results

---

🏗️ Project Structure

FaceGCD/
├── data_loader/              # Dataset loaders and augmentations
│   ├── augmentations/        # Data augmentation strategies
│   ├── youtube_faces_*.py    # YTF dataset loaders
│   ├── casia_webface_*.py    # CASIA dataset loaders
│   └── data_loaders.py       # Main data loading utilities
├── model/                    # Model architectures
│   ├── dino_vision_transformer.py  # DINO ViT backbone
│   ├── prefix_generator.py   # HyperNetwork-based prefix generator
│   ├── ViT_face.py           # Face-specific ViT components
│   └── mobilenet.py          # Landmark CNN
├── trainer/                  # Training utilities
│   ├── trainer.py            # Main training loop
│   └── faster_mix_k_means_pytorch.py  # Semi-supervised K-Means
├── utils/                    # Utility functions
│   ├── cluster_utils.py      # Clustering utilities
│   ├── losses.py             # Loss functions
│   └── dino_utils.py         # DINO-specific utilities
├── shell/                    # Shell scripts for experiments
├── train.py                  # Training script
├── extract_features.py       # Feature extraction script
├── SSK.py                    # Semi-supervised K-Means evaluation
└── requirements.txt          # Python dependencies

---

📝 Citation

If you find this work useful for your research, please cite:

@article{oh2025facegcd,
  title={FaceGCD: Generalized Face Discovery via Dynamic Prefix Generation},
  author={Oh, Yunseok and Choi, Dong-Wan},
  journal={arXiv preprint arXiv:2507.22353},
  year={2025}
}

---

🙏 Acknowledgments

This work is built upon several excellent projects:

• DINO - Self-supervised Vision Transformers
• GCD - Generalized Category Discovery
• ArcFace - Face Recognition baseline
• timm - PyTorch Image Models
• APEX - Mixed Precision Training

---

📧 Contact

For questions or issues, please:

• Open an issue on GitHub
• Contact: oys5339@inha.edu

---

📜 License

This project is released under the MIT License. See LICENSE file for details.

---

🔄 Updates

• [2025-10] Initial release of code and paper
• [Coming Soon] Pretrained models and datasets will be available via Google Drive

---

Note: This repository is actively maintained. Dataset and checkpoint download links will be updated once the upload to Google Drive is complete.