Sphereface face recognition + Intel OpenVINO



The repository contains the entire pipeline (including all the preprocessings) for deep face recognition with SphereFace. The recognition pipeline contains three major steps: face detection, face alignment and face recognition.

SphereFace is a recently proposed face recognition method. It was initially described in an arXiv technical report and then published in CVPR 2017. The most up-to-date paper with more experiments can be found at arXiv or here. To facilitate the face recognition research, we give an example of training on CAISA-WebFace and testing on LFW using the 20-layer CNN architecture described in the paper (i.e. SphereFace-20).

In SphereFace, our network architecures use residual units as building blocks, but are quite different from the standrad ResNets (e.g., BatchNorm is not used, the prelu replaces the relu, different initializations, etc). We proposed 4-layer, 20-layer, 36-layer and 64-layer architectures for face recognition (details can be found in the paper and prototxt files). We provided the 20-layer architecure as an example here. If our proposed architectures also help your research, please consider to cite our paper.

SphereFace achieves the state-of-the-art verification performance (previously No.1) in MegaFace Challenge under the small training set protocol.

StartDT-AI 4/16/2018 93.8226% Large
StartDT-AI we only use a single model trained on a ResNet-28 network joined with cosine loss and triplet loss on MS-Celeb-1M(74K identities, 4M images), refer to

DeepVisage: Making face recognition simple yet with powerful generalization skills


One-shot Face Recognition by Promoting Underrepresented Classes


ArcFace: Additive Angular Margin Loss for Deep Face Recognition


SphereFace: Deep Hypersphere Embedding for Face Recognition



Algorithm Date Submitted Set 1 Set 2 Set 3 Data Set Size
Sogou AIGROUP – SFace 9/5/2018 99.939% 99.939% 99.939% Large
Sogou We collected and filtered millions pictures from sogou pic search engine, and also used pictures date cleaned by DeepInsight. We trained multi model by different loss functions such as combined margin loss, margin loss with focal loss and so on. In some models we divided faces into different patches so we can get features represent local feature such as mouth and even teeth. We merge the features get from different models together as the final result.

Leave a Reply

Your email address will not be published. Required fields are marked *