- version 0.4 updating;
There may be some bugs in the code,please contact me if you find some of them,thank you~
Redo ArcFace Additive Angular Margin Loss for Deep Face Recognition，some of re-releazing codes are referenced and list below,appreciate so much to your opensource.
- https://github.com/deepinsight/insightface (Original)
- https://github.com/luckycallor/InsightFace-tensorflow (Much Appreciate)
- https://github.com/auroua/InsightFace_TF
- https://github.com/tensorflow/models
version 0.4:
- ubuntu16.04 + 2*GTX 1080ti + Python3.6 + Anaconda5.2.0 + Tensorflow1.7-gpu + MySQL5.7.25 + PyQt5
| model | lfw | calfw | cplfw | agedb_30 | cfp_ff | cfp_fp | lfw_face |
|---|---|---|---|---|---|---|---|
| resnet_v2_m_50(prelu) | 0.979 | 0.878 | 0.823 | 0.886 | 0.976 | 0.904 | 0.866 |
| resnet_v2_m_50(leaky_relu) | 0.992 | 0.932 | 0.860 | 0.935 | 0.990 | 0.910 | 0.943 |
lfw_face is the lfw dataset that generate by my owm MTCNN model
- Face recognization accuracy is greatly disturbed by face detection;
- One of the most important and confuse things is to select the proper threshold to recognize;
Face recognizer is in /recognizer/arcface_recognizer.py，Face detection is based on my MTCNN repository；
overview of face-recognizer's interface：
Basic interface：
- get_embd ： interface to get face embeddings，shape=[n,512];
- align_face ： interface of face align;
- recognize ： interface of face recognization, output recognization result and bounding-box(es) of face(s);
Extended interface：
- add_customs ： add fresh person data to database;
- add_embds ： update embedding data of target member(s) from database(old data will update by average increment);
- update_customs ： update embedding data of target member(s) from database(old data will be replaced);
- del_customs : delete data of target member(s) from database;
It's a basic GUI ,a more functional interface will be updated in future version.
