Head Pose Estimation

[takuya-takeuchi]

From Matrix

Reference

https://github.com/lincolnhard/head-pose-estimation

How to

• Implement as C++ code on FaceRecognitionDotNet.Native.HeadPoseEstimation
  • Import OpenCV
  • Not use OpenCVSharp to reduce depending on 3rd party library
• Simple. We can get results from landmark data.

From Deep Learning

Reference

https://github.com/guozhongluo/head-pose-estimation-and-face-landmark

How to

• Implement as C++ code on DlibDotNet.Native.Dnn.HeadPoseEstimation
• Not simple. Because use face image data and pass to deep neural network!!

From SVM

Reference

https://sukohi.blogspot.com/2017/03/scikit-learn-head-pose-estimation.html

How to

• Maybe, we use only dlib. And all C#

Memo

• https://github.com/natanielruiz/deep-head-pose/blob/master/code/datasets.py

[takuya-takeuchi]

Memo

https://github.com/guozhongluo/head-pose-estimation-and-face-landmark/blob/d6b5816727565de9031b77b2d267c5f8567de22d/model/deploy.prototxt#L194

As far as I can see this, it needs only lanrmark point to estimate pose? If so, it may be able to simplify network.

[takuya-takeuchi]

ops, poselayer code is missing...

[takuya-takeuchi]

poselayer is InnerProduct

[takuya-takeuchi]

WIP.

• Multilayer perceptron (use loss_multiclass_log not mlp)
• input 68 points not face image

[takuya-takeuchi]

       Roll: 13.3179502210815
      Pitch: -25.4978731056757
        Yaw: -46.5588758170786
 Predicted Roll: 9.62651096074842
Predicted Pitch: -16.836335144937
  Predicted Yaw: -48.8857434364036

GT

Predicted

[takuya-takeuchi]

Use KRLS instead of Dnn. It works fine.

GT

Predicted

[takuya-takeuchi]

Implemented HeadPoseEstimator. Get Face Landmark by FRDN and estimate head pose.

GT

Predicted (all manual)