人脸识别方案汇总

 

• SphereFace: Deep Hypersphere Embedding for Face Recognition(A-softloss)

	MegaFace	LFW(Labeled Faces in the Wild)	YouTube Faces DB
Face Verification	89.142%(#3)	99.42%(#6)	95.0%(#8)
Face Identification	75.766%(#3)	/

• CosFace:Large Margin Cosine Loss for Deep Face Recognition

	MegaFace	LFW(Labeled Faces in the Wild)	YouTube Faces DB
Face Verification	96.65%(# 2)	99.73%(# 2)	97.6%(# 3)
Face Identification	82.72%(# 2)	/	/

• ArcFace: Additive Angular Margin Loss for Deep Face Recognition

	MegaFace	LFW(Labeled Faces in the Wild)	YouTube Faces DB
Face Verification	98.48%,(# 1) (model:ArcFace + MS1MV2 + R100,)	99.83%,(# 1) (model: ArcFace + MS1MV2 + R100,)	98.02%,(# 2) (Model: ArcFace + MS1MV2 + R100,)
Face Identification	98.35%,(# 1) (model: ArcFace + MS1MV2 + R100 + R)	/	/

• Additive Margin Softmax for Face Verification（AM-softmax）

实验数据较少，但在ResNet 20中比A-softmax中比A-softmax要好，got 97.96 on LFW under FAR=0.1% 

• Support Vector Guided Softmax Loss for Face Recognition

	MegaFace	LFW(Labeled Faces in the Wild)	Trillion Pairs
Face Verification	99.03 SV-AM-Softmax	97.57% SV-Arc-Softmax	78.49% SV-AM-Softmax loss
Face Identification	98.82 SV-AM-Softmax	97.20% SV-AM-Softmax	82.25% SV-AM-Softmax loss