maybe someday it goes to work out my questions~
2018/08/25 Now we come a new stage and I should learn more professional knowledge.
What I have learnt from
This tells you how to normally setup
https://blog.csdn.net/jia20003/article/details/54583431
and when you got some trouble：Pls check your network(My home network trapped me for a long while)
Problems such as
Download face_landmark_model.dat Failed (or time out)
missing ffmpeg
but the download is real slow. you could use your mobile and provide hotspot for your PC.(As I done)
if auto-download can't do to help, you could check these issues :
https://github.com/opencv/opencv_contrib/issues
At last ,google can be your best driver.
Finally, got it! After check and check again, it works!
Below is the code of printing all files' name in current catalog
Just paste it in the PowerShell
Attention : the F:\OCV\opencv\newbuild\install\x64\vc15\lib is my file directory path
Get-ChildItem F:\OCV\opencv\newbuild\install\x64\vc15\lib | ForEach-Object -Process{
if($_ -is [System.IO.FileInfo])
{
Write-Host($_.name);
}
}
In the OCV3 Project I don't use src as default input image but use img_1.Which leads me misleading and make some mistakes. In the last days, I should take care of this!!
These two days, i just watch corner detection, one of detection methods is Harris , and another is Shi-Tomasi Corner detetion. In my point of view , Good Features To Track (aka Shi-Tomasi) performs better than Harris detection .
But both of them did not mark the top of roof which human could easily recognize.
Pic blow is good feature to track
and till now, I should have finished custom corner detection, but I think it is unnecessary to learn. When I meet such kind of projects or problems, I would come back have a careful seek.
These two days I have tried SURF and SIFT,both of them are using to detect KeyPoints in the image which is hard for human beings to recognize what it is.
And the result of Experiments is that there are not so much differences between SURF and SIFT, but you still say, the KeyPoints of image shows their own features, which we can conclude that SIFT seems better?
Almost forget to say, their sample code looks same.
//SIFT跟SURF代码是一模一样的
//numOfFeatures指的是特征点的个数
int numOfFeatures = 400;
//现在创建检测器
Ptr<SIFT> detector = SIFT::create(numOfFeatures);
vector<KeyPoint> keypoints;//存到这来
//检测
detector->detect(src, keypoints);
Mat kpImage;
//绘制关键点
drawKeypoints(src, keypoints, kpImage);
namedWindow("result", WINDOW_AUTOSIZE);
imshow("result", kpImage);HOG Descriptor
Here is how to calculate size of HOG pls refer this Link https://blog.csdn.net/yangpan011/article/details/78353873
And in the section was tought how to use HOG to establish "Descriptor" detector and use SVM method to detect how many people in the picture you provide.
Just like this.
Obviously, it is not precise to detect all people in the pic.
Maybe next time I will learn how to fix this issue.
Here is code
//SVM 检测人群 7938000个描述子 速度较慢
HOGDescriptor hog = HOGDescriptor();
hog.setSVMDetector(hog.getDefaultPeopleDetector());
vector<Rect> foLocations;
hog.detectMultiScale(src, foLocations,0, Size(8, 8), Size(32, 32), 1.05);
Mat newSrc = src.clone();
for (size_t i = 0; i < foLocations.size(); i++)
{
rectangle(newSrc, foLocations[i], Scalar(0, 255, 0),3);
}
imshow("hhh", newSrc);Yesterday, we learnt how to use HOG to recoginize the count of people. And today we are practicing to use descriptors to match a thing in another pic such as looking for a book in the gallery.
Picture blow shows what we are looking for.
And this is the pic we download form Internet.
After compiling these code:
//作用：匹配两张图像
//需要两个描述子
//本次采用SURF描述子
Ptr<SURF> detector = SURF::create(400);
//储存两个描述子的keypoint
vector<KeyPoint> keyPoint_1;
vector<KeyPoint> keyPoint_2;
//声明两个描述子
Mat descriptor_1, descriptor_2;
detector->detectAndCompute(src, Mat(), keyPoint_1, descriptor_1);
detector->detectAndCompute(img_1, Mat(), keyPoint_2, descriptor_2);
//匹配
BFMatcher bfMatcher;
vector<DMatch> matches;
bfMatcher.match(descriptor_1, descriptor_2, matches);
//绘画
Mat resImg;
drawMatches(src, keyPoint_1, img_1, keyPoint_2, matches, resImg);
imshow("res", resImg);We got :
The steps of code and why I write this show in comments.
Using API FlannBasedMatcher, I got the same result as BruteForce Match.
The filter is necessary to use, and the effect is obviously.You could esaily recognize them.
However I could not totally understand how it works.
Ptr<KAZE> detector = KAZE::create();
vector<KeyPoint> keyPoints;
detector->detect(src, keyPoints);
printf("KAZE");
Mat keyPointImage;
drawKeypoints(src, keyPoints, keyPointImage, Scalar::all(-1), DrawMatchesFlags::DEFAULT);;
imshow("kaze res", keyPointImage);As the description said, the performance is better than SURF and SHIF, but I could not tell.
Here is the KAZE image:
AKAZE accelerates the process time of KAZE, just replace "KAZE" by "AKAZE". Other codes keep same.
BRISK took a short time.And the "Face Detection" really annoyed me.
After fuking two months learning ,just a few lines of code could be implement. I am regreting why I learning for so long.
Fine,it is ok. And time for now I should write how to use Face Detection Code:
Here is fucking fish code：
First we should find our OpenCV install folder where we could find
And full path that I used in code.which you could replace yours:
String caPath = "F:/OCV/opencv/newbuild/install/etc/haarcascades/haarcascade_frontalface_alt.xml";Ok,then you could just paste my code and press F5 run the program.
#pragma region FaceDetection 人脸检测
//在Surface上需要更改位置（用的是自带的face训练集）
String caPath = "F:/OCV/opencv/newbuild/install/etc/haarcascades/haarcascade_frontalface_alt.xml";
CascadeClassifier face_cascade;
if (!face_cascade.load(caPath))
{
printf("face cascade could not load");
waitKey(0);
return -1;
}
cvtColor(src, graySrc, COLOR_BGR2GRAY);
equalizeHist(graySrc, graySrc);
vector<Rect> faces;
face_cascade.detectMultiScale(graySrc, faces,1.1,2,0,Size(30,30));
for (size_t i = 0; i < faces.size(); i++)
{
rectangle(src, faces[i], Scalar(0, 0, 255));
}
namedWindow("face detection", WINDOW_AUTOSIZE);
imshow("face detection", src);
#pragma endregion
the result shows:
Oh, btw,you need to replace your face picture path,which named "src".And graySrc is temporary Mat to store gray pic.
代码可参考： https://blog.csdn.net/weixin_41695564/article/details/80077706
And here is my own experinence : https://segmentfault.com/a/1190000017221248
In another way, this article should be read carefully : https://www.cnblogs.com/skyfsm/p/6879265.html which tells how to write code of boder detection.
The full experience is recorded at here : https://segmentfault.com/a/1190000017257827
First of all, we should access camera to record one or more persons faces.And in the code, While(cap.read(frame)) is used to save capture image,but if not give your programe a pause or delay ,there may be some problems occour.
here is what I have done:
-
for Static Pictures
This pic is an openning book on the desk.
and a Surface default BG
the function of K-Means I still have not understand...
-
for Real-time video capture
it is the time I am in Lib
recently I am thinking of how to choose between ASP.Net and OpenCV. My company needs me to learn ASP.Net to build some websites, which is aimed to optimize workflow.
My teacher told me that solving your working issue which leads your direction.OpenCV now could be added to wait list.When I get free time, I may learn it again.
These days I am too lazy to update, but from now on. I will keep code updating!!!