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