We have trained three deep networks (ResNet-101, ResNet-152, ResNet-200) with joint softmax and triplet loss on MS-Celeb-1M (95K identities, 5.1M images), and the triplet part is trained by batch online hard negative mining with subspace learning. The features of all networks are concatenated to produce the final feature, whose dimension is set to be 256x3. For data processing, we use original large images and follow our own system by detection and alignment. Particularly, in evaluation, we have cleaned the FaceScrub and MegaFace with the code released by iBUG_DeepInsight.

Compared to our another submission named “Orion Star Technology (clean)”, the major difference is that no any data cleaning is adopted in evaluation.

This is a model ensembled by three different models using ResNet CNN and improved ResNet network, learned by a combined loss. A filtered MS-Celeb-1M and CASIA-Webface is used as the dataset.

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

Large-Margin Softmax Loss for Convolutional Neural Networks https://arxiv.org/abs/1612.02295

A Discriminative Deep Feature Learning Approach for Face Recognition

https://ydwen.github.io/papers/WenECCV16.pdf

NormFace: L2 Hypersphere Embedding for Face Verification

https://arxiv.org/abs/1704.06369

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

Adopt the ensemble of very deep CNNs, learned by joint supervision (softmax loss, improved center loss, etc). Training data is a combination of public datasets (CAISA, VGG, CACD2000, etc) and private datasets. The total number of images is more than 2 million.

Adopt a network of very deep ResNet CNNs, learned by combined supervision(identification loss(softmax loss), verification loss, triplet loss).

The algorithm consists of three parts: FCN - based fast face detection algorithm, pre-training ResNet CNN on classification task, weight tuning. Training set contains 273000 photos. Hardware: 8 x Tesla k80.

SphereFace uses a novel approach to learn face features that are discriminative on a hypersphere manifold. The training data set we use in SphereFace is the publicly available CASIA-WebFace dataset which contains 490k images of nearly 10,500 individuals.

arcface, https://github.com/deepinsight/insightface

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

https://arxiv.org/abs/1703.08388

One-shot Face Recognition by Promoting Underrepresented Classes

https://arxiv.org/abs/1707.05574v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Compares to our submissions named “BingMMLab V1(cleaned data)”, the only difference is that no data cleaning is adopted in this evaluation.

We used knowledge graph to collect identities and then crawled Bing search engine to get high quality images, we filtered noises in 14M training data by clustering with weak face models, and then trained a DensetNet-69 (k=48) network with A-Softmax loss variants. In evaluation, we used the cleaned test set released by iBUG_DeepInsight.

DenseNet

https://arxiv.org/pdf/1608.06993v2.pdf

A-Softmax and its variant:

https://arxiv.org/pdf/1704.08063.pdf

https://github.com/wy1iu/LargeMargin_Softmax_Loss/issues/13

Angular Softmax Loss with Channel-wise Attention

We have trained our model on ResNet-152 with Additive Angular Margin Loss on combined dataset with MS-Celeb-1M and VggFace2, and cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

VGGFace2: A dataset for recognising faces across pose and age

https://arxiv.org/abs/1710.08092

MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition

https://www.microsoft.com/en-us/research/wp-content/uploads/2016/08/MSCeleb-1M-a.pdf

We have trained ResNet models with a combined loss on MS-Celeb-1M. In evaluation, we have cleaned the FaceScrub and MegaFace with the code released by iBUG_DeepInsight.

[iBUG_DeepInsight code](https://github.com/deepinsight/insightface)

Compares to our submissions named “TUPUTECH v1 (clean)”, the only difference is data cleaning by TUPU is adopted in evaluation.

trained network with arcface loss. tried some different methods in preprocessing data.In evaluation, used the cleaned test set released by iBUG_DeepInsight.

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

MTCNN:Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

FaceNet: A Unified Embedding for Face Recognition and Clustering

https://arxiv.org/abs/1503.03832

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

MTCNN:Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

FaceNet: A Unified Embedding for Face Recognition and Clustering

https://arxiv.org/abs/1503.03832

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

we only use a single model trained on a ResNet-101 network joined with cosine loss and triplet loss on MS-Celeb-1M(74K identities, 4M images), refer to

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

A Discriminative Deep Feature Learning Approach for Face Recognition

https://ydwen.github.io/papers/WenECCV16.pdf

using insightface with loss modification.

Custom version of ArcFace.

FeelingFace, which trained with cleaned MS-Celeb-1M dataset

We have trained our model on ResNet-152 with Additive Angular Margin Loss on combined dataset with MS-Celeb-1M and VggFace2, and cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

VGGFace2: A dataset for recognising faces across pose and age

https://arxiv.org/abs/1710.08092

MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition

https://www.microsoft.com/en-us/research/wp-content/uploads/2016/08/MSCeleb-1M-a.pdf

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.

We trained the deep residual attention network(attention-56) with A-softmax to learn the face feature.

The training dataset is constructed by the novel dataset building techinique, which is critical for us to improve the performance of the model. The results are the cleaned test set performance released by iBUG_DeepInsight.

Wang F, Jiang M, Qian C, et al. Residual attention network for image classification[J]. arXiv preprint arXiv:1704.06904, 2017.

Deng J, Guo J, Zafeiriou S. ArcFace: Additive Angular Margin Loss for Deep Face Recognition[J]. arXiv preprint arXiv:1801.07698, 2018.

Liu W, Wen Y, Yu Z, et al. Sphereface: Deep hypersphere embedding for face recognition[C]//The IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017, 1.

arcface, https://github.com/deepinsight/insightface

A single deep Resnet model (the 'r100' configuration from insightface) trained with our own angular margin loss (an improved variant of A-Softmax, not published yet).

The training set consists of nearly 6.2M images (about 96K identities). For MegaFace evaluation, we adopted the clean list released by iBUG_DeepInsight.

insightface: https://github.com/deepinsight/insightface

A-Softmax: https://arxiv.org/pdf/1704.08063.pdf

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

This is a single model trained by a deep ResNet network on MS-Celeb-1M，learned by a cosine loss, refer to ArcFace: Additive Angular Margin Loss for Deep Face Recognition https://arxiv.org/abs/1801.07698 SphereFace: Deep Hypersphere Embedding for Face Recognition https://arxiv.org/abs/1704.08063 We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

mtcnn: Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

https://github.com/pangyupo/mxnet_mtcnn_face_detection

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

This feature model is an ensemble of 3 deep Resnet models (the 'r100' and 'r152' configuration from insightface[1] ) trained with our own angular margin loss (an improved variant of A-Softmax[2] , not published yet).

The training set consists of 7 Million images (about 188K identities). For evaluation, we adopted the clean list released by iBUG_DeepInsight[1] .

Reference:

[1] insightface: https://github.com/deepinsight/insightface

[2] A-Softmax: https://arxiv.org/pdf/1704.08063.pdf

A single model (improved Resnet-152) is trained by the supervision of combined loss functions (A-Softmax loss, center loss, triplet loss et al) on MS-Celeb-1M (84 k identities, 5.2 M images). In evaluation, we use the cleaned FaceScrub and MegaFace released by iBUG_DeepInsight.