Kazemi-based facemark detection pre-trained model does not produce valid fitting results

[cedriclmenard]

System information (version)

• OpenCV => 3.4.1
• Operating System / Platform => MacOS 10.13.5
• Compiler => LLVM (using CMake)

Detailed description

The pre-trained Kazemi facemark detection model does not produce valid results when using OpenCV's own face detection (i.e. Haar's cascade using the LBP improved trained model).

Essentially, the landmarks always seem to be off or at a completely other place than what it should. I'm experiencing the same issue as: https://stackoverflow.com/questions/48080914/what-model-is-needed-to-detect-facemark-by-using-opencv-contrib

The facemark model file is available here: https://github.com/opencv/opencv_3rdparty/tree/contrib_face_alignment_20170818

The face detection model xml is available here: https://github.com/opencv/opencv/blob/master/data/lbpcascades/lbpcascade_frontalface_improved.xml

[krshrimali]

Hi!

Yes, the pre-trained model did not show valid results for me. Can you share some details on your code?

1. Sample test image and output result
2. The code used for testing (there may be some small logical errors, so just need to cross-check)