MobileFaceNets: Efficient CNNs for Accurate Real-time Face Verification on Mobile Devices

By Sheng Chen, Yang Liu, Xiang Gao, Zhen Han. Chinese Conference on Biometric Recognition, 2018

706 citations recorded by the source corpus.

We present a class of extremely efficient CNN models, MobileFaceNets, which use less than 1 million parameters and are specifically tailored for high-accuracy real-time face verification on mobile and embedded devices. We first make a simple analysis on the weakness of common mobile networks for face verification. The weakness has been well overcome by our specifically designed MobileFaceNets. Under the same experimental conditions, our MobileFaceNets achieve significantly superior accuracy as well as more than 2 times actual speedup over MobileNetV2. After trained by ArcFace loss on the refined MS-Celeb-1M, our single MobileFaceNet of 4.0MB size achieves 99.55% accuracy on LFW and 92.59% TAR@FAR1e-6 on MegaFace, which is even comparable to state-of-the-art big CNN models of hundreds MB size. The fastest one of MobileFaceNets has an actual inference time of 18 milliseconds on a mobile phone. For face verification, MobileFaceNets achieve significantly improved efficiency over previous state-of-the-art mobile CNNs.

Subfields: Computer Vision and Pattern Recognition, Signal Processing

View Paper: https://www.semanticscholar.org/paper/d7cbedbee06293e78661335c7dd9059c70143a28

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

1. Jie Zhou
2. Jiwen Lu
3. Xiaogang Wang
4. Luc Van Gool
5. Chen Change Loy
6. Dacheng Tao
7. Yu Liu
8. Yi Yang

In Signal Processing

1. Jie Zhou
2. Hang Zhou
3. Yu Liu
4. Jiwen Lu
5. Guanglu Song
6. Manyuan Zhang
7. Feng Zhou
8. Qiang Meng

Citations in the corpus

• VarGFaceNet: An Efficient Variable Group Convolutional Neural Network for Lightweight Face Recognition (2019, methodology): “To validate the efficiency and effectiveness of VarGFaceNet, we first train our network from scratch, and compare the performance with mobilefacenet(y2) [2, 3].”
• OneFlow: Redesign the Distributed Deep Learning Framework from Scratch (2021, methodology): “observe that OneFlow’s throughput slightly outperforms InsightFace’s when training face recognition models with ResNet and MobileFaceNet as backbone networks respectively (Chen et al., 2018).”
• ShrinkTeaNet: Million-scale Lightweight Face Recognition via Shrinking Teacher-Student Networks (2019, positive): “For all the experiments, we use the Resnet-90 structure [4] as the teacher network, while different light-weight networks, i.e. MobileNetV1 [32], MobileNetV2 [33], MobileFaceNet [34].”
• Discriminability Distillation in Group Representation Learning (2020, methodology): “As for the base model, we also experiment DDL with MobileFaceNet [6], a popular backbone for mobile devices.”
• Learning Compatible Embeddings (2021, methodology): “The old model is trained with ResNet50 while the new models are trained with ResNet100, ResNet18 [15] or MobileFaceNet [5].”
• Efficient Acceleration of Deep Learning Inference on Resource-Constrained Edge Devices: A Review (2023, methodology): “In [385], a CNN named Mobile-FaceNets is presented that uses a global depthwise convolution filter that can perform face recognition having less than 1M parameters.”
• WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition (2021, methodology): “Due to strict time limitation, models constrained by FRUITS-100 can only adopt lightweight architectures, including RetinaFace-MobileNet-0.”
• CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition (2020, neutral): “, MobileFaceNet [6].”
• Dynamic Region-Aware Convolution (2020, methodology): “We use MobileFaceNet [3] as our backbone which has only 1M parameters and 189M MADDs with input size 112× 96.”
• End-to-End Learning of Energy-Constrained Deep Neural Networks (2018, methodology): “For optimizing W , we use SGD with a learning rate η1 = 0.01 on AlexNet, LeNet-5, and MobileNetV2, and use η1 = 0.001 on SqueezeNet.”