some face recognition models based on GPU/CPU(intergrated with face detection and alignment)
| Baseline Model | Accuracy(LFW) | GPU\CPU | Framework | Detection | alignment |
|---|---|---|---|---|---|
| Facenet | 99.65% | GPU | Tensorflow | MTCNN | MTCNN |
| Openface | 99.63%(in paper) 92.92% | both | dlib+opencv, C++ | dlib | dlib |
| Center Face(ResNet) | 99.03% | gpu | caffe | no | no |
| normface | used for improving | gpu | caffe | MTCNN | MTCNN |
| seetaface | 97.1% | cpu | no | yes | yes |
| dlib | 99.38% | cpu | no | dlib | dlib |
Both Facenet and openface are the implementation of paper facenet_2015. In openface, the real accuracy is lower because of the bad detection of dlib.
If you are under computation-limited environment(cpu only or embedded system), seetaface(v1 and v2) and dlib are prefered. Although they can not provide best accuracy, the running speed is fast enough.
If you have GPUs, the Facenet are recommanded. Beacause it is highly acdamical and technical supported.