pntt3011 / mediapipe_face_iris_cpp

Real-time Face and Iris Landmarks Detection using C++

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Do you have mediapipe_hand_landmark_cpp ?

lexidliny opened this issue · comments

HI @lexidliny, i have a separate branch for it. https://github.com/pntt3011/mediapipe_face_iris_cpp/tree/HandLandmark. Hope this help.

Thank you ! I have tested this code on my computer and I get "Segmentation fault (core dumped)". How do I fix this fault?
The face_landmark_cpp program works well on my computer with ubuntu 18.04.

Can you show your detailed error? I tested on my computer and it ran normally.
Did you compile Debug or Release mode? If you compile with Debug mode, try again with Release mode.

I have tried again, and I located this error in the my::DetectionPostProcess::getHighestScoreDetection function.

@lexidliny, did you change the define values in DetectionPostProcess.hpp? They are different from those of Face landmark model.

I found that I forgot to modfy the path to anchors.csv. Now it works well. Thank you !

It seems that the accuracy is not as good as mediapipe. Is it caused by palm_detection_without_custom_layer.tflite ?

It seems that the accuracy is not as good as mediapipe. Is it caused by palm_detection_without_custom_layer.tflite ?

@lexidliny, if i remember correctly, the landmark model works best with straight up hand.
Therefore, from the detection model, you should:

  • Retrieve the 5 (or 6?) main landmarks
  • Calculate the rotate angle from the straight up direction
  • Rotate the hand bounding box the opposite of that angle
  • Crop the bounding box
    Then you can feed it to the landmark model.

You can get the straight up direction from palm and middle finger landmark (2 of the output landmarks from detection model)