Year | Publication | Contribution | Network architecture | ILSVRC error (top-5) (%) | Face recognition system |
---|---|---|---|---|---|
2012 | AlexNet [34] | Large DCNN with 60M parameters and 650,000 neurons | Ensemble of 7 models | 15.3 | - |
2014 | VGGNet [35] | Increasing depth using very small convolution filters. | Ensemble of 2 models | 6.8 | VGGFace [37] |
DeepId3 [28] | |||||
2014 | GoogleNet [36] | Inception architecture | Ensemble of 7 models | 6.67 | FaceNet [29] |
DeepId3 [28] | |||||
OpenFace [65] | |||||
2014 | InceptionV2[73] | Adding batch normalization to Inception architecture | Batch Normalized Inception ensemble | 4.9 | - |
2014 | InceptionV3 [74] | Adding factorization to Inception architecture | Ensemble of four Inception-V3s | 3.58 | - |
2015 | Residual Learning [70] | A residual learning framework to ease the training of very deep CNNs | ResNet 34 | 5.60 | DLIB [64] |
ResNet 50 | 5.25 | ArcFace [38] | |||
ResNet 101 | 4.60 | CosFace [61] | |||
ResNet 152 | 4.49 | SphereFace [60] | |||
Ensemble | 3.57 | - | |||
2016 | InceptionV4 [71] | Adding residual learning on top of Inception | Inception-ResNet-v1 | 4.3 | FaceNet_Re [66] |
Inception-ResNet-v2 | 3.7 | of | |||
Ensemble of 4 DNNs | 3.1 |