Face Verification on LFW (test)
99.87Verification AccuracyLeaderboard 1st Place
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Leaderboard 1st PlaceNote=External baseline2018.12 | 99.87 | — | |
| TopoFRTraining Data=Glint360K, Backbone=R200, Venue=NeurIPS242024.10 | 99.87 | — | |
| ResNet200-TopoFRYear='24, Param (M)=118.8, FLOPs (M)=23.5G, Training Dataset=Glint360K, Latency (ms)=12.242026.04 | 99.87 | — | |
| SV-AM-SoftmaxBackbone=Attention-56 [31] with IRSE [4], Training Dataset=MS-Celeb-1Mv1c2018.12 | 99.85 | — | |
| GroupFaceBackbone=ResNet101, Training Dataset=MS1MV22021.05 | 99.85 | — | |
| BroadFaceTraining Data=MS1MV2, Backbone=R100, Venue=ECCV202024.10 | 99.85 | — | |
| TopoFRTraining Data=MS1MV2, Backbone=R100, Venue=NeurIPS242024.10 | 99.85 | — | |
| TopoFR†Training Data=MS1MV2, Backbone=R100, Venue=NeurIPS242024.10 | 99.85 | — | |
| TopoFR†Training Data=MS1MV2, Backbone=R200, Venue=NeurIPS242024.10 | 99.85 | — | |
| TopoFRTraining Data=MS1MV2, Backbone=R200, Venue=NeurIPS242024.10 | 99.85 | — | |
| TopoFRTraining Data=Glint360K, Backbone=R50, Venue=NeurIPS242024.10 | 99.85 | — | |
| TransFace-BTraining Data=Glint360K, Backbone=ViT, Venue=ICCV232024.10 | 99.85 | — | |
| TopoFRTraining Data=Glint360K, Backbone=R100, Venue=NeurIPS242024.10 | 99.85 | — | |
| TransFace-LTraining Data=Glint360K, Backbone=ViT, Venue=ICCV232024.10 | 99.85 | — | |
| TransFace-LYear='25, Param (M)=271.6, FLOPs (M)=25.4G, Training Dataset=Glint360K, Latency (ms)=OOM2026.04 | 99.85 | — | |
| TransFace-BYear='25, Param (M)=124.5, FLOPs (M)=11.5G, Training Dataset=Glint360K, Latency (ms)=18.202026.04 | 99.85 | — | |
| TransFace-SYear='23, Param (M)=86.7, FLOPs (M)=5.8G, Training Dataset=Glint360K, Latency (ms)=14.312026.04 | 99.85 | — | |
| Transformer-CosFaceTrain Data=MS1MV2, Backbone=ResNet100, Embedding Size=5122024.12 | 99.83 | — | |
| ArcFaceTrain Data=MS1MV2, Backbone=ResNet100, Embedding Size=5122024.12 | 99.83 | — | |
| Transformer-ArcFaceTrain Data=MS1MV2, Backbone=ResNet100, Embedding Size=5122024.12 | 99.83 | — | |
| ArcFaceTrain Data=WebFace4M, Backbone=ResNet100, Embedding Size=5122024.12 | 99.83 | — | |
| ArcFaceModels=1 ResNet-100, Data=MS-Celeb-1M2018.03 | 99.83 | — | |
| SeqFaceModels=1 ResNet-64, Data=MS-Celeb-1M + Celeb-Seq2018.03 | 99.83 | — | |
| PartialFC-r1.0Backbone=ResNet101, Training Dataset=MS1MV2, Sampling Rate=1.02021.05 | 99.83 | — | |
| ArcFace (LResNet100E-IR)FLOPs (M)=24211, Parameters (M)=65.22021.07 | 99.83 | — | |
| ArcFace#Images=5.8M, Backbone=R100, Training Data=MS1MV22022.05 | 99.83 | — | |
| MagFace*Backbone=ResNet-1002022.03 | 99.83 | — | |
| CosFaceBackbone=ResNet-1002022.03 | 99.83 | — | |
| ArcFaceBackbone=ResNet-1002022.03 | 99.83 | — | |
| Arc+UNPGBackbone=ResNet-1002022.03 | 99.83 | — | |
| TopoFR†Training Data=MS1MV2, Backbone=R50, Venue=NeurIPS242024.10 | 99.83 | — | |
| TopoFRTraining Data=MS1MV2, Backbone=R50, Venue=NeurIPS242024.10 | 99.83 | — | |
| MagFace+Training Data=MS1MV2, Backbone=R100, Venue=CVPR212024.10 | 99.83 | — | |
| AdaFaceTraining Data=MS1MV2, Backbone=R200, Venue=CVPR222024.10 | 99.83 | — | |
| TransFace-LTraining Data=MS1MV2, Backbone=ViT, Venue=ICCV232024.10 | 99.83 | — | |
| AdaFaceTraining Data=Glint360K, Backbone=R200, Venue=CVPR222024.10 | 99.83 | — | |
| Transformer-AdafaceTrain Data=WebFace4M, Backbone=ResNet100, Embedding Size=5122024.12 | 99.82 | — | |
| Transformer-CosfaceTrain Data=WebFace4M, Backbone=ResNet100, Embedding Size=5122024.12 | 99.82 | — | |
| Transformer-ArcfaceTrain Data=WebFace4M, Backbone=ResNet100, Embedding Size=5122024.12 | 99.82 | — | |
| ArcFaceBackbone=ResNet101, Training Dataset=MS1MV22021.05 | 99.82 | — | |
| PartialFC-r0.1Backbone=ResNet101, Training Dataset=MS1MV2, Sampling Rate=0.12021.05 | 99.82 | — | |
| SFace#Images=5.8M, Backbone=R100, Training Data=MS1MV22022.05 | 99.82 | — | |
| MagFace2022.07 | 99.82 | — | |
| ArcFace*Backbone=ResNet-1002022.03 | 99.82 | — | |
| AdaFaceTraining Data=MS1MV2, Backbone=R50, Venue=CVPR222024.10 | 99.82 | — | |
| SCF-ArcFaceTraining Data=MS1MV2, Backbone=R100, Venue=CVPR212024.10 | 99.82 | — | |
| AdaFaceTraining Data=MS1MV2, Backbone=R100, Venue=CVPR222024.10 | 99.82 | — | |
| TransFace-BTraining Data=MS1MV2, Backbone=ViT, Venue=ICCV232024.10 | 99.82 | — | |
| AdaFaceTraining Data=Glint360K, Backbone=R50, Venue=CVPR222024.10 | 99.82 | — | |
| AdaFaceTraining Data=Glint360K, Backbone=R100, Venue=CVPR222024.10 | 99.82 | — | |
| ArcFaceTraining Data=Glint360K, Backbone=R200, Venue=CVPR192024.10 | 99.82 | — | |
| CosFaceTrain Data=MS1MV2, Backbone=ResNet100, Embedding Size=5122024.12 | 99.81 | — | |
| Cos+UNPGBackbone=ResNet-1002022.03 | 99.81 | — | |
| MagFaceBackbone=ResNet-1002022.03 | 99.81 | — | |
| Mag+UNPGBackbone=ResNet-1002022.03 | 99.81 | — | |
| ArcFaceTraining Data=Glint360K, Backbone=R100, Venue=CVPR192024.10 | 99.81 | — | |
| AdaFaceTrain Data=WebFace4M, Backbone=ResNet100, Embedding Size=5122024.12 | 99.8 | — | |
| CosFaceTrain Data=WebFace4M, Backbone=ResNet100, Embedding Size=5122024.12 | 99.8 | — | |
| SeqFaceModels=1 ResNet-27, Data=MS-Celeb-1M + Celeb-Seq2018.03 | 99.8 | — | |
| CurricularFaceBackbone=ResNet101, Training Dataset=MS1MV22021.05 | 99.8 | — | |
| DCQBackbone=ResNet101, Training Dataset=MS1MV22021.05 | 99.8 | — | |
| CurricularFace2022.07 | 99.8 | — | |
| ElasticFace2022.07 | 99.8 | — | |
| MV-SoftmaxTraining Data=MS1MV2, Backbone=R100, Venue=AAAI202024.10 | 99.8 | — | |
| CurricularFaceTraining Data=MS1MV2, Backbone=R100, Venue=CVPR202024.10 | 99.8 | — | |
| ElasticFace-Cos+Training Data=MS1MV2, Backbone=R100, Venue=CVPR222024.10 | 99.8 | — | |
| FaceLiVTv2-LParam (M)=8.52, FLOPs (M)=309, Training Dataset=Glint360K, Latency (ms)=0.71(17.2× ↓)2026.04 | 99.8 | — | |
| ArcFaceTraining Data=MS1MV2, Backbone=R200, Venue=CVPR192024.10 | 99.79 | — | |
| ArcFaceModels=1 ResNet-50, Data=MS-Celeb-1M2018.03 | 99.78 | — | |
| CosFace (ours)Backbone=ResNet101, Training Dataset=MS1MV22021.05 | 99.78 | — | |
| CosFaceTraining Data=MS1MV2, Backbone=R100, Venue=CVPR182024.10 | 99.78 | — | |
| URLTraining Data=MS1MV2, Backbone=R100, Venue=CVPR202024.10 | 99.78 | — | |
| ArcFaceTraining Data=Glint360K, Backbone=R50, Venue=CVPR192024.10 | 99.78 | — | |
| ResNet50-ArcFaceYear='22, Param (M)=43.6, FLOPs (M)=6.3G, Training Dataset=Glint360K, Latency (ms)=3.762026.04 | 99.78 | — | |
| FaceLiVTv2-MParam (M)=7.02, FLOPs (M)=258, Training Dataset=Glint360K, Latency (ms)=0.65(18.8× ↓)2026.04 | 99.78 | — | |
| FaceLiVTv2-SParam (M)=4.62, FLOPs (M)=179, Training Dataset=Glint360K, Latency (ms)=0.54(22.7× ↓)2026.04 | 99.78 | — | |
| BaiduModels=9, Data=1.2M images2018.03 | 99.77 | — | |
| ArcFaceTraining Data=MS1MV2, Backbone=R100, Venue=CVPR192024.10 | 99.77 | — | |
| MagFaceTraining Data=MS1MV2, Backbone=R50, Venue=CVPR212024.10 | 99.74 | — | |
| CosFaceModels=1 ResNet-64, Data=5M images2018.03 | 99.73 | — | |
| CosFace#Images=5M2022.05 | 99.73 | — | |
| Circle-loss*Backbone=ResNet-1002022.03 | 99.73 | — | |
| Mobiface#Image=3.8M, Training epoch=1024, Params=2.4M, MAdds=462M, Batch size=10242019.08 | 99.7 | — | |
| Seesaw-shuffleFaceNet#Image=5.8M, Training epoch=16, Params=1.3M, MAdds=146M, Batch size=160, Training dataset=MS1MV22019.08 | 99.7 | — | |
| MobileFaceNetFLOPs (M)=933, Parameters (M)=22021.07 | 99.7 | — | |
| VarGFaceNetFLOPs (M)=1022, Parameters (M)=52021.07 | 99.68 | — | |
| MixFaceNet-MFLOPs (M)=626.1, Parameters (M)=3.952021.07 | 99.68 | — | |
| ArcFaceTraining Data=MS1MV2, Backbone=R50, Venue=CVPR192024.10 | 99.68 | — | |
| ShuffleFaceNet 1.5xFLOPs (M)=577.5, Parameters (M)=2.62021.07 | 99.67 | — | |
| Tencent-BestImage#Nets=20, Training Set (private or Public face dataset)=1,000, 000 images of 20, 000 subjects, Private, Training Setting=Joint-Bayes2015.07 | 99.65 | 0.25 | |
| FaceNetModels=1, Data=200M images2018.03 | 99.65 | — | |
| Seesaw-shuffleFaceNet(mobi)#Image=5.8M, Training epoch=16, Params=2.8M, MAdds=154M, Batch size=224, Training dataset=MS1MV22019.08 | 99.65 | — | |
| FaceNet#Nets=1, Training Set (private or Public face dataset)=100 ~ 200 million images of 8 million subjects, Private, Training Setting=L22015.07 | 99.63 | 0.09 | |
| FaceNet#Images=200M2022.05 | 99.63 | — | |
| FaceLiVTv2-XSParam (M)=2.9, FLOPs (M)=90, Training Dataset=Glint360K, Latency (ms)=0.43(28.5× ↓)2026.04 | 99.63 | — | |
| ShuffleFaceNet 2xFLOPs (M)=1050, Parameters (M)=4.52021.07 | 99.62 | — | |
| Seesaw-shuffleFaceNet#Image=5.8M, Training epoch=16, Params=1.3M, MAdds=146M, Batch size=192, Training dataset=MS1MV22019.08 | 99.6 | — | |
| ShuffleMixFaceNet-MFLOPs (M)=626.1, Parameters (M)=3.952021.07 | 99.6 | — | |
| MixFaceNet-SFLOPs (M)=451.7, Parameters (M)=3.072021.07 | 99.6 | — | |
| MixFaceNet-XSFLOPs (M)=161.9, Parameters (M)=1.042021.07 | 99.6 | — |