分享一个针对移动端的人脸算法全栈项目，包含的模块会比较全，会陆续更新人脸的全套技术，包括检测、关键点、矫正、姿态、属性、识别、人脸品质等等，对学习人脸很有帮助哦~

作者：becauseofAI
项目地址：becauseofAI/MobileFace
项目还在持续更新中，欢迎大家下载试用、提issue 、点star 以资鼓励

MobileFace

A face recognition solution on mobile device.

Prerequirements

• Anaconda (optional but recommend)
• MXNet and GluonCV (the easiest way to install)
• DLib (may be deprecated in the future)
  The easiest way to install DLib is through pip.

  pip install dlib

Performance

Identification

Model Framework Size CPU LFW Target MobileFace_Identification_V1 MXNet 3.40M 8.5ms - Actual Scene MobileFace_Identification_V2 MXNet 3.41M 9ms 99.653% Benchmark MobileFace_Identification_V3 MXNet 2.10M 3ms(sota) 95.466%(baseline) Benchmark

Detection

Model Framework Size CPU MobileFace_Detection_V1 MXNet/GluonCV 30M 20ms/50fps

Landmark

Model Framework Size CPU MobileFace_Landmark_V1 DLib 5.7M <1ms

Example

To get fast face feature embedding with MXNet as follow:

cd example
python get_face_feature_mxnet.py

To get fast face detection result with MXNet/GluonCV as follow:

cd example
python get_face_boxes_gluoncv.py

To get fast face landmarks result with dlib as follow:

cd example
python get_face_landmark_dlib.py

Visualization

t-SNE

I used the t-SNE algorithm to visualize in two dimensions the 256-dimensional embedding space. Every color corresponds to a different person(but colors are reused): as you can see, the MobileFace has learned to group those pictures quite tightly. (the distances between clusters are meaningless when using the t-SNE algorithm)

To get the t-SNE feature visualization above as follow:

cd tool/tSNE
python face2feature.py # get features and lables and save them to txt
python tSNE_feature_visualization.py # load the txt to visualize face feature in 2D with tSNE

ConfusionMatrix

I used the ConfusionMatrix to visualize the 256-dimensional feature similarity heatmap of the LFW-Aligned-100Pair: as you can see, the MobileFace has learned to get higher similarity when calculating the same person's different two face photos. Although the performance of the V1 version is not particularly stunning on LFW Dataset, it does not mean that it does not apply to the actual scene.

To get the ConfusionMatrix feature similarity heatmap visualization above as follow:

cd tool/ConfusionMatrix
python ConfusionMatrix_similarity_visualization.py

To get inference time of different version's MXNet models as follow:

cd tool/time
python inference_time_evaluation_mxnet.py --symbol_version=V3 # default = V1

Model_Prune

Prune the MXNet model through deleting the needless layers (such as classify layer and loss layer) and only retaining features layers to decrease the model size for inference as follow:

cd tool/prune
python model_prune_mxnet.py

MXNet2Caffe

Merge_bn

Benchmark

The LFW test dataset (aligned by MTCNN and cropped to 112x112) can be download from Dropbox or BaiduDrive, and then put it (named lfw.bin) in the directory of data/LFW-bin.
To get the LFW comparison result and plot the ROC curves as follow:

cd benchmark/LFW
python lfw_comparison_and_plot_roc.py

MegaFace

• [x] MobileFace_Identification
• [x] MobileFace_Detection
• [x] MobileFace_Landmark
• [ ] MobileFace_Align
• [ ] MobileFace_Attribute
• [ ] MobileFace_Pose
• [ ] MobileFace_NCNN
• [ ] MobileFace_FeatherCNN
• [x] Benchmark_LFW
• [ ] Benchmark_MegaFace

Others

Coming Soon!

Reference

微信公众号: 极市平台（ID: extrememart ）
每天推送最新CV干货