Person detector from MediaPipe Pose

[LaurentBerger]

Ok that's not in the list

[WanliZhong]

👍👍👍Thanks for your contribution! Here are some code need to be modified.

[WanliZhong]

LGTM

[LaurentBerger]

LGTM

please take a look at https://forum.opencv.org/t/get-same-results-in-c-and-python-code/13656

[fengyuentau]

please take a look at https://forum.opencv.org/t/get-same-results-in-c-and-python-code/13656

Is it a key factor in this pull request? The only place where convertTo is used is converting landmarks to int32 at the final stage.

---

convertTo uses saturate_cast which follows the rounding rule, but astype from numpy just truncates anything behind the floating dot.

You can loop over all landmarks and use something like the following for a truncation convertion:

int landmark_x = static_cast<int>(results.at<float>(row, 4);

[LaurentBerger]

In this PR there is only one convertTo but I have finished to translate pose_estimation_mediapipe/demo.py and change are visble in image result. It is like reading same file using videocpature or imread. input file is same but inference is not equal.
MAy be it is a detail but when you want to check result you look for an error that's does not exist

[LaurentBerger]

don't push there is a bug