Enhancement: Accept Grayscale as Input

[EzequielAdrianM]

Currently when detection is performed, the code reads the channels of a Mat object as input. But if it is in grayscale (Channels == 1) it is converted again to color:

detector.h

virtual inline int det(cv::Mat& image) {
    if(image.empty()) return 0;
    if(image.channels() == 1) {
        cv::cvtColor(image, image, CV_GRAY2BGR);
    } CHECK(image.channels() == 3);
    dlib::cv_image<dlib::bgr_pixel> img(image);
    mRets = mFaceDetector(img);
    return mRets.size();
}

However, the Dlib documentation points out that: "Any built in scalar type may be used as a grayscale pixel type. For example, unsigned char, int, double, etc."
So bgr_pixel type is not the only alternative.

[EzequielAdrianM]

Solution:

virtual inline int det(cv::Mat& image) {
  if(image.empty()) return 0;
  if(image.channels() > 1) {
    //Convert to GrayScale
    cv::cvtColor(image, image, CV_BGR2GRAY);
  }
  dlib::cv_image<uchar> img(image);
  mRets = mFaceDetector(img);
  return mRets.size();
}

This is really speeding up the detection!!! Hope helping somebody

[e-snail]

@EzequielAdrianM Thanks!

cv::cvtColor(image, image, CV_BGR2GRAY); leads to an error. Instead, I try following code in jni/jni_detections/jni_face_det.cpp :

cv::cvtColor(rgbaMat, bgrMat, cv::COLOR_RGBA2GRAY); //replace cv::cvtColor(rgbaMat, bgrMat, cv::COLOR_RGBA2BGR);

But it did NOT work in speeding up the detection. Any idea ?

[EzequielAdrianM]

It does not work because you are not sending the gray scaled mat to the detector. You keep sending the original in color.
cv::cvtColor(bgrMat, grayMat, cv::COLOR_BGR2GRAY);
Here, "grayMat" is the new simplified Mat you must give to the detector. But you keep sending "image" (the original full color Mat):
dlib::cv_image<unsigned char> img(grayMat);

[e-snail]

@EzequielAdrianM I did send gray scaled mat to the detector, NOT the bgr mat. For ur convenience, I adjust the code sequence as following:

in jni/jni_detections/detector.h

cv::Mat image_gray;
if (image.channels() > 1) {
cv::cvtColor(image, image_gray, CV_BGR2GRAY);
}
//CHECK(image.channels() == 3);
// TODO : Convert to gray image to speed up detection
// It's unnecessary to use color image for face/landmark detection
dlib::cv_image<unsigned char> img(image_gray);
mRets = mFaceDetector(img);

Is it right ?

[EzequielAdrianM]

Yes, it is right