[論文レビュー] Face.evoLVe: A High-Performance Face Recognition Library

Qingzhong Wang, Pengfei Zhang+2 more|arXiv (Cornell University)|Jul 19, 2021

Face recognition and analysis参考文献 96被引用数 27

ひとこと要約

Face.evoLVeは、PyTorchとPaddlePaddle全体にわたる高性能顔認識のための包括的でモジュール式のライブラリで、多くのバックボーン、損失関数、データのテクニック、事前学習済みモデルを提供し、公正で再現性のある比較を可能にします。

ABSTRACT

In this paper, we develop face.evoLVe -- a comprehensive library that collects and implements a wide range of popular deep learning-based methods for face recognition. First of all, face.evoLVe is composed of key components that cover the full process of face analytics, including face alignment, data processing, various backbones, losses, and alternatives with bags of tricks for improving performance. Later, face.evoLVe supports multi-GPU training on top of different deep learning platforms, such as PyTorch and PaddlePaddle, which facilitates researchers to work on both large-scale datasets with millions of images and low-shot counterparts with limited well-annotated data. More importantly, along with face.evoLVe, images before & after alignment in the common benchmark datasets are released with source codes and trained models provided. All these efforts lower the technical burdens in reproducing the existing methods for comparison, while users of our library could focus on developing advanced approaches more efficiently. Last but not least, face.evoLVe is well designed and vibrantly evolving, so that new face recognition approaches can be easily plugged into our framework. Note that we have used face.evoLVe to participate in a number of face recognition competitions and secured the first place. The version that supports PyTorch is publicly available at https://github.com/ZhaoJ9014/face.evoLVe.PyTorch and the PaddlePaddle version is available at https://github.com/ZhaoJ9014/face.evoLVe.PyTorch/tree/master/paddle. Face.evoLVe has been widely used for face analytics, receiving 2.4K stars and 622 forks.

研究の動機と目的

• Provide a unified, reproducible framework for state-of-the-art face recognition methods.
• Cover the full face recognition pipeline: detection, alignment, feature extraction, and loss heads.
• Support multiple platforms (PyTorch and PaddlePaddle) with distributed multi-GPU training.
• Offer a modular design enabling easy plugging of custom backbones and heads.
• Release preprocessed data, trained models, and training/evaluation environments to facilitate fair benchmarking.

提案手法

• Modular pipeline: face detection, landmark-based alignment, backbone feature extraction, and loss head computation.
• Support for multiple backbones (e.g., ResNet variants, IR, LightCNN, MobileNet, GhostNet, EfficientNet, etc.).
• A wide range of loss functions (e.g., Softmax, ArcFace, CosFace, SphereFace, Triplet, ArcNegFace, CircleLoss, MagFace, etc.) with optional focal variants.
• Bag of tricks for training stability and performance (warmup, cosine annealing, label smoothing, distillation/knowledge transfer, curricular approaches).
• Data processing steps include handling low-shot classes, augmentation, balanced sampling, and dataset-wide preprocessing for reproducibility.
• Cross-platform training support with distributed training to scale on large datasets (e.g., MS-Celeb-1M, WebFace260M).
• Provision of pre-processed data, trained models, and evaluation pipelines for fair benchmarking.

実験結果

リサーチクエスチョン

• RQ1 How can a unified, modular library accelerate reproducible research in face recognition across different backbones and losses?
• RQ2 What is the impact of diverse backbones and angular/loss functions on standardized face recognition benchmarks under fair settings?
• RQ3 Can distributed training on PyTorch and PaddlePaddle effectively scale to large-scale datasets while maintaining fair comparisons?
• RQ4 Do training tricks (warmup, cosine annealing, label smoothing, distillation) meaningfully improve stability and accuracy across models?
• RQ5 How do released pre-processed datasets and trained models facilitate reproducibility and rapid experimentation?

主な発見

Backbone	Head	Loss	Testing Dataset	LFW	CFP_FF	CFP_FP	AgeDB	CALFW	CPLFW	Vggface2_FP
IR152	Arcface	Focal	LFW	99.82	99.84	98.37	98.07	96.03	93.05	95.50
IR101	Arcface	Focal	LFW	99.81	99.74	98.25	97.77	95.93	92.74	95.44
IR50	AdaCos	Focal	LFW	99.75	99.53	98.39	97.25	95.55	92.25	95.27
IR152	AdaCos	Focal	LFW	99.82	99.84	98.37	98.07	96.03	93.05	95.50
HRNet	MV-Softmax	Focal	LFW	99.82	99.51	98.41	97.88	95.43	88.95	94.70
TF-NAS-A	AM-Softmax	Focal	LFW	99.82	99.47	98.33	96.65	94.32	84.88	91.38
GhostNet	ArcFace	Focal	LFW	99.69	99.52	98.48	97.29	94.92	85.25	90.88
AttentionNet	AdaCos	Focal	LFW	99.82	99.47	98.52	96.89	95.12	87.23	94.23
MobileFaceNet	AdaCos	Focal	LFW	99.73	99.84	97.75	95.87	94.87	89.29	93.20

• The library supports numerous backbones (e.g., IR50/101/152, ResNet variants, LightCNN, MobileNet, GhostNet, EfficientNet) and losses (ArcFace, CosFace, SphereFace, AM-Softmax, AdaCos, CircleLoss, MagFace, etc.).
• Tables show competitive accuracy across standard benchmarks (LFW, CALFW, CPLFW, AgeDB, CFP variants, VggFace2) using various backbone/head/loss combinations.
• IR152 with AdaCos or ArcFace losses achieves high results on LFW, CALFW, and CPLFW when trained on large-scale data (MS-Celeb-1M or Web260M).
• Face.evoLVe demonstrates competitive or superior performance compared to other toolboxes (e.g., FaceX-zoo) on key datasets.
• Distributed training is supported, enabling scaling to datasets with millions of images.
• The library has achieved state-of-the-art results in open competitions and continues to evolve with active contributors.