netscope of this network is here:prototxt, visualization
mode file is here:baidu pan
| LFW | CFP | AgeDB30 |
|---|---|---|
| 99.8 | 98.7 | 97.0 |
update:
2018/11/2:train by insight_face_emore,the model file is here:baidu pan
some motivation of this network:
1.the main calculation of ResNet is focus on the convolution on front layer
2.the main accuracy part of ResNet is the 13 Residual construction operation on 14x14 feature map
3.feature Embedding of Invert Residual Construction can automaticlly choose the most discriminative feature.