First, make sure you have OPENCV and MXNET already installed.

• OPENCV How to install OPENCV.

  ​ We use cv2 for image IO and resize(much faster than skimage), the input image's channel is acutally BGR

• MXNET How to install MXNET.

  ​ please update to the newest version, we need 'full' mode in Pooling operation

  ​ Only tested on Linux and Windows.

We use a Robust Initialization for Cascaded Pose Regression (RICPR) by providing texture and pose correlated initial shapes for the testing face.

MTCNN:

Test

cd initialization/mxnet_mtcnn_face_detection/

Run:

python main.py

Epnp_Mapping

Run the following file:

initialization/Epnp_Mapping/ConsoleApplication1/ConsoleApplication1.cpp

Run the following file:

initialization/Texture_initialization/getlbpHist.m

we use a global cascaded pose regression with robust initialization to get localization results for the original image

Run the following file:

rcpr_with_initialization/FULL_demoRCPR.m

Run the following file:

rcpr_with_initialization_mirror/FULL_demoRCPR.m

Run the following file:

correction_model/Correlation_model/zhenshi_jiance.m
correction_model/Correlation_model/jiance_jingxiang.m
correction_model/Correlation_model/myPearson.m

The global average correlation of mirror error and localization error on the COFW dataset:

correction_model/dingquezhi_tichu.m
correction_model/zhengli.m

we use a global cascaded pose regression with robust initialization to get localization results for the original image

Run the following file:

Local/rcpr_v1_stable_left_eye/FULL_demoRCPR.m
Local/rcpr_v1_stable_right_eye/FULL_demoRCPR.m
Local/rcpr_v1_stable_nose/FULL_demoRCPR.m
Local/rcpr_v1_stable_mouse/FULL_demoRCPR.m

Run the following file:

Local/Integration.m

Run the following file:

Local_mirror/rcpr_v1_stable_left_eye/FULL_demoRCPR.m
Local_mirror/rcpr_v1_stable_right_eye/FULL_demoRCPR.m
Local_mirror/rcpr_v1_stable_nose/FULL_demoRCPR.m
Local_mirror/rcpr_v1_stable_mouse/FULL_demoRCPR.m

Run the following file:

Local_mirror/Integration.m

Run the following file:

correction_model/Correlation_model/zhenshi_jiance.m
correction_model/Correlation_model/jiance_jingxiang.m
correction_model/Correlation_model/myPearson.m

The local average correlation of mirror error and localization error on the COFW dataset：

correction_model/dingquezhi_tichu.m
correction_model/zhengli.m

In the following figure, the red markers in the rst row is unreliable results and the green markers in the first row is reliable results. The second row is the adjusted results on the HELEN dataset.

Here is the results and some examples on COFW datasets. More results is shown in the folder "results".

Results on COFW datasets:

Examples of facial landmark localization comparison on COFW. The rows from top to button are the examples of RCPR, HPM, RICPR and TSCPR(Ours), respectively.

More examples on COFW datasets: