Accelerating Face Recognition Technology Development: Open Source Libraries and Datasets

Face recognition is ubiquitous in our lives, for example, in building access control systems, where it replaces traditional access cards or passwords, enhancing convenience and security. In terms of mall security, face recognition is widely used in monitoring systems, helping to identify and track potential criminals or missing persons, thus improving safety measures. Additionally, unlocking smartphones is another important application of face recognition technology, providing users with a quick and convenient method of identity verification, replacing traditional passwords or fingerprint recognition.

Face recognition technology involves processing and analyzing large-scale image and video data. To ensure accurate and efficient face recognition, substantial computational resources must be invested to support key processes such as image processing, feature extraction, model training, and inference. Therefore, as face recognition technology evolves and becomes more widely used, the demand for computational resources continues to grow.

This article aims to introduce several open-source libraries and datasets for face recognition to help developers accelerate research and application in this field.

DeepFace

DeepFace is a lightweight Python framework for face recognition and facial attribute analysis (age, gender, emotion, and ethnicity). It is a hybrid face recognition framework that includes state-of-the-art models: VGG-Face, Google FaceNet, OpenFace, Facebook DeepFace, DeepID, ArcFace, Dlib, and SFace.

Main features: face detection, face alignment, feature extraction, face verification, face search, face clustering, face attribute recognition, face tracking, facial expression recognition, ethnicity recognition, gender recognition, etc.

• Open source address: https://github.com/serengil/deepface

OpenFace

OpenFace is an open-source face recognition and verification library developed by Carnegie Mellon University (CMU). It can perform facial landmark detection, head pose estimation, facial action unit recognition, eye-gaze estimation, and facial feature extraction tasks.

• Open source address: https://github.com/TadasBaltrusaitis/OpenFace

FaceNet

FaceNet is a deep learning-based facial recognition system developed by Google. It is a multipurpose recognition system that can be used simultaneously for face verification (whether it’s the same person), identification (who this person is), and clustering (finding similar people).

• Open source address: https://github.com/davidsandberg/facenet
• Paper address: https://arxiv.org/abs/1503.03832

InsightFace

InsightFace is an open-source 2D and 3D deep face analysis toolbox, primarily based on PyTorch and MXNet. It effectively implements a rich variety of face recognition, face detection, and face alignment, optimized for training and deployment.

• Open source address: https://github.com/deepinsight/insightface

CelebA

CelebA is an open-source facial attribute dataset released by The Chinese University of Hong Kong, widely used for face-related computer vision tasks, including face attribute identification, face detection, and landmark marking. This dataset contains 20,259 images of 10,177 celebrity identities, each image annotated with feature labels, including face bounding box annotations, coordinates of 5 facial feature points, and 40 attribute labels.

Subsequently, several related datasets based on CelebA have also been released: CelebA-Dialog, CelebAMask-HQ, and CelebA-Spoof. Among them, CelebA-Dialog is a large-scale visual-language face dataset with rich fine-grained labels, categorizing an attribute into multiple levels based on its semantics; CelebAMask-HQ consists of 30,000 high-resolution facial images selected from the CelebA dataset, each image has a corresponding segmentation mask for CelebA’s facial attributes. The mask size of CelebAMask-HQ is 512 × 512 and includes 19 types of attribute features, including skin, nose, eyes, eyebrows, ears, mouth, lips, hair, hats, glasses, earrings, necklaces, neck, fabric, and other facial parts and decorative accessories; CelebA-Spoof is a face liveness detection dataset containing 625,537 images of 10,177 subjects, with 43 rich attributes related to face, lighting, environment, and deception types. Out of the 43 attributes, 40 pertain to live images, encompassing all facial information such as skin, nose, eyes, eyebrows, lips, hair, hats, and glasses; 3 attributes pertain to spoof images, including deception type, environment, and lighting conditions.

• Download address: https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

MegaFace

MegaFace is a large-scale public face recognition training dataset, considered one of the most important benchmarks for commercial face recognition vendors. This dataset includes 4,753,320 faces of 672,057 identities from 3,311,471 photos downloaded from the albums of 48,383 Flickr users.

• Download address: https://megaface.cs.washington.edu/dataset/download.html

CASIA-CeFA

CASIA-CeFA is a public face liveness detection dataset with ethnic labels, containing 1,607 subjects across 3 ethnicities and 3 modes, as well as 2D and 3D attack types.

• Paper address: https://arxiv.org/abs/2003.05136

Glint360K

Glint360K is claimed to be the largest and cleanest face recognition dataset in the world, open-sourced by DeepGlint, containing 18 million images of 360,000 categories.

• Download address: https://github.com/deepinsight/insightface/tree/master/recognition/partial_fc#glint360k

As a computational service provider, Qudong Cloud possesses high-performance computational resources capable of rapidly processing massive data, providing strong support for face recognition algorithms. Furthermore, Qudong Cloud has thousands of datasets, including face recognition-related datasets such as AR-Face-Database and CelebAMask-HQ. These datasets are available for developers to use with a single click, greatly facilitating research and application in face recognition technology.