This library provides a comprehensive suite for high-performance face recognition, targeting researchers and engineers. It offers tools for face alignment, data processing, and training state-of-the-art face recognition models using various backbones and loss functions, enabling rapid development and deployment of face analytics applications.

How It Works

The library leverages deep learning architectures, supporting numerous backbone networks (e.g., ResNet, IR-SE) and loss functions (e.g., ArcFace, CosFace, Focal Loss). It features an advanced distributed training schema that scales to large identity sets by parallelizing both the backbone and the final classification layer across multiple GPUs, overcoming limitations of single-master approaches.

Quick Start & Requirements

• Install: Clone the repository.
• Prerequisites: Linux/macOS, Python 3.7+, PyTorch 1.0+, OpenCV 3.4.5+. CUDA-enabled GPU recommended (4-8 NVIDIA Tesla P40 used in development). Optional dependencies include MXNet, TensorFlow, and tensorboardX.
• Setup: Requires data preparation in a specific directory structure.
• Docs: Homepage

Highlighted Details

• Supports a wide range of backbone architectures and loss functions for flexible model development.
• Includes a robust distributed training strategy for large-scale face recognition.
• Provides extensive data processing utilities, including alignment and handling imbalanced datasets.
• Offers pre-trained models and detailed performance benchmarks on various public datasets (LFW, CFP, AgeDB, etc.).

Maintenance & Community

The project was developed at Tencent FiT DeepSea AI Lab and is associated with the Learning and Vision (LV) group at the National University of Singapore. The README indicates a merger with Baidu PaddlePaddle in 2021.

Licensing & Compatibility

The code is released under the MIT License, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

The README mentions that some datasets require direct contact for access due to licensing issues. The training script requires specific data directory structures and configuration file setup.