aligning the face

[MyraBaba]

Hi, @Qengineering

what I noticed that you are aligning the face in TWrap:

 cv::warpPerspective(SmallFrame, aligned, m, cv::Size(96, 112), cv::INTER_LINEAR);
  resize(aligned, aligned, cv::Size(112, 112), 0, 0, cv::INTER_LINEAR);

I saw from ncnn repo:
[
cv::warpAffine(img_src.clone(), *face_aligned, transfer_mat, cv::Size(112, 112), 1, 0, 0);)

is this make a huge diference ? you are usingb (cv::Size(96, 112)) after that resing , below one ise directly 112,112

would you mind to look at it ?

[Qengineering]

The difference is subtle.
In warpAffine() is *face_aligned a [2x3] matrix, as where in warpPerspective, aligned a [3x3] matirx is. For more info see the OpenCV docs.

[MyraBaba]

For some faces which is almost +90 degree turned sometimes alignment make face upside down

…

On 26 Dec 2023, at 13:03, Q-engineering ***@***.***> wrote: The difference is subtle. In warpAffine() is *face_aligned a [2x3] matrix, as where in warpPerspective, aligned a [3x3] matirx is. For more info see the OpenCV docs <https://docs.opencv.org/3.4/da/d6e/tutorial_py_geometric_transformations.html#:~:text=warpAffine%20takes%20a%202x3%20transformation,3x3%20transformation%20matrix%20as%20input.>. — Reply to this email directly, view it on GitHub <#19 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEFRZH3O2TPQEIE4CHUNJPTYLKOGJAVCNFSM6AAAAABBDC4IC2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNRZGQZDKNRQGI>. You are receiving this because you authored the thread.

[Qengineering]

If not implemented, try to get a check on the rotation before acceptance.

[MyraBaba]

@Qengineering

here after aligment :

[Qengineering]

I see.
The key to solve this issue is the aligned [3x3] matirx.
The factors determine how the image is transformed.
Investigate the factors. It is relatively simple to detect a large rotation.
The OpenCV docs give you more informations.

[MyraBaba]

Hi, In Pythonn vesion I never came cross such upside down rotation (180 degree) PS: what you recommend as a model for rpi zero w 2 for human detection ? Try to be below 35usd cost wise Best

…

On 27 Dec 2023, at 12:46, Q-engineering ***@***.***> wrote: I see. The key to solve this issue is the aligned [3x3] matirx. The factors determine how the image is transformed. Investigate the factors. It is relatively simple to detect a large rotation. The OpenCV docs <https://docs.opencv.org/3.4/da/d6e/tutorial_py_geometric_transformations.html#:~:text=warpAffine%20takes%20a%202x3%20transformation,3x3%20transformation%20matrix%20as%20input.> give you more informations. — Reply to this email directly, view it on GitHub <#19 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEFRZHYEA7Y33J7JF5CSYKLYLPU5ZAVCNFSM6AAAAABBDC4IC2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNZQGEZTANRZGQ>. You are receiving this because you authored the thread.

[MyraBaba]

@Qengineering
would you mind to give me the example cpp part ? really confused on c++

[Qengineering]

Tip:

You: Can you calculate in a C++ program using the OpenCV function cv::warpPerspective(SmallFrame, aligned, m, cv::Size(96, 112), cv::INTER_LINEAR); a rotation angle of the image in degrees?

Answer ChatGTP:
Certainly! To calculate the rotation angle from a perspective transformation matrix, you can use the following C++ code. This assumes that the transformation matrix m is a 3x3 matrix obtained from the cv::getPerspectiveTransform function or some other method.

#include <opencv2/opencv.hpp>

int main() {
    // Read the original image
    cv::Mat SmallFrame = cv::imread("path/to/your/image.jpg");

    // Define the transformation matrix (you should have this matrix from somewhere)
    cv::Mat m = (cv::Mat_<double>(3, 3) << 1, 0, 0, 0, 1, 0, 0, 0, 1);

    // Perform the perspective transformation
    cv::Mat aligned;
    cv::warpPerspective(SmallFrame, aligned, m, cv::Size(96, 112), cv::INTER_LINEAR);

    // Calculate rotation angle
    double angle = atan2(m.at<double>(1, 0), m.at<double>(0, 0)) * 180.0 / CV_PI;

    // Print the rotation angle
    std::cout << "Rotation angle: " << angle << " degrees." << std::endl;

    return 0;
}

In this code, atan2 is used to calculate the angle from the transformation matrix elements. The rotation angle is stored in the variable angle and then printed. Note that this assumes the transformation matrix represents only a rotation and not a more complex transformation. If your transformation matrix includes other elements, you may need a more sophisticated approach to extract rotation information.

[MyraBaba]

Where are we heading.. These AI … Or humanity is soooo low intellectuality so we are impressed by AI ? Thnx

…

On 27 Dec 2023, at 20:17, Q-engineering ***@***.***> wrote: Tip: You: Can you calculate in a C++ program using the OpenCV function cv::warpPerspective(SmallFrame, aligned, m, cv::Size(96, 112), cv::INTER_LINEAR); a rotation angle of the image in degrees? Answer ChatGTP: Certainly! To calculate the rotation angle from a perspective transformation matrix, you can use the following C++ code. This assumes that the transformation matrix m is a 3x3 matrix obtained from the cv::getPerspectiveTransform function or some other method. #include <opencv2/opencv.hpp> int main() { // Read the original image cv::Mat SmallFrame = cv::imread("path/to/your/image.jpg"); // Define the transformation matrix (you should have this matrix from somewhere) cv::Mat m = (cv::Mat_<double>(3, 3) << 1, 0, 0, 0, 1, 0, 0, 0, 1); // Perform the perspective transformation cv::Mat aligned; cv::warpPerspective(SmallFrame, aligned, m, cv::Size(96, 112), cv::INTER_LINEAR); // Calculate rotation angle double angle = atan2(m.at<double>(1, 0), m.at<double>(0, 0)) * 180.0 / CV_PI; // Print the rotation angle std::cout << "Rotation angle: " << angle << " degrees." << std::endl; return 0; } In this code, atan2 is used to calculate the angle from the transformation matrix elements. The rotation angle is stored in the variable angle and then printed. Note that this assumes the transformation matrix represents only a rotation and not a more complex transformation. If your transformation matrix includes other elements, you may need a more sophisticated approach to extract rotation information. — Reply to this email directly, view it on GitHub <#19 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEFRZH4RXE7EMYD5NLVLU7TYLRJYZAVCNFSM6AAAAABBDC4IC2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNZQGQ4DQMJZGE>. You are receiving this because you authored the thread.

[Qengineering]

Believe me, I have the same feelings towards AI as you. Hesitated a lot.
However, a free clever 'student' at my side, finding some answers, which I undoubtedly had found also, but a few times faster, enlighten my rather heavy workload.
I'm not impressed by AI. It's boring when you look at close.

[MyraBaba]

@Qengineering

Quick note:

double angle = atan2(m.at<double>(1, 0), m.at<double>(0, 0)) * 180.0 / CV_PI;

is not producing right result unless you change to double to float.

double angle = atan2(m.at<float>(1, 0), m.at<float>(0, 0)) * 180.0 / CV_PI;

by the way I try to use yolov5 cpp face detection (the version which has not accuracy percentage return) .
what would be the superior over scrfd and yolov5 face detection ? do you have any experience for yolov7/8 cpp face detection ?
Best

[Qengineering]

I haven't any experience of YoloV7 or V8 with face detection.
Please note that not only the faces need to be detected, but also the position of the eyes, nose and mouth.
I'm not sure if the face detectors based on Yolo feature this.

[MyraBaba]

Yolov5 face detector also has landmarks with it 5 point.

…

On 28 Dec 2023, at 22:22, Q-engineering ***@***.***> wrote: I haven't any experience of YoloV7 or V8 with face detection. Please note that not only the faces need to be detected, but also the position of the eyes, nose and mouth. I'm not sure if the face detectors based on Yolo feature this. — Reply to this email directly, view it on GitHub <#19 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEFRZHZBIX7R4KUXGE25NDLYLXBH7AVCNFSM6AAAAABBDC4IC2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNZRGQZTGOJQGE>. You are receiving this because you authored the thread.

[Qengineering]

OK.. You could try it. However, YoloV5 might be a lot slower. (??)

[MyraBaba]

What is your suggestion ? Retinaface ?

…

On 28 Dec 2023, at 23:28, Q-engineering ***@***.***> wrote: OK.. You could try it. However, YoloV5 might be a lot slower. (??) — Reply to this email directly, view it on GitHub <#19 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEFRZH3IA7KEO5IB6VWOO4LYLXI53AVCNFSM6AAAAABBDC4IC2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNZRGQ3TAMBUHE>. You are receiving this because you authored the thread.