I have a folder of positives and another of negatives images in JPG format, and I want to train an SVM based on that images, I've done the following but I receive an error:
Mat classes = new Mat();
Mat trainingData = new Mat();
Mat trainingImages = new Mat();
Mat trainingLabels = new Mat();
CvSVM clasificador;
for (File file : new File(path + "positives/").listFiles()) {
Mat img = Highgui.imread(file.getAbsolutePath());
img.reshape(1, 1);
trainingImages.push_back(img);
trainingLabels.push_back(Mat.ones(new Size(1, 1), CvType.CV_32FC1));
}
for (File file : new File(path + "negatives/").listFiles()) {
Mat img = Highgui.imread(file.getAbsolutePath());
img.reshape(1, 1);
trainingImages.push_back(img);
trainingLabels.push_back(Mat.zeros(new Size(1, 1), CvType.CV_32FC1));
}
trainingImages.copyTo(trainingData);
trainingData.convertTo(trainingData, CvType.CV_32FC1);
trainingLabels.copyTo(classes);
CvSVMParams params = new CvSVMParams();
params.set_kernel_type(CvSVM.LINEAR);
clasificador = new CvSVM(trainingData, classes, new Mat(), new Mat(), params);
When I try to run that I obtain:
OpenCV Error: Bad argument (train data must be floating-point matrix) in cvCheckTrainData, file ..\..\..\src\opencv\modules\ml\src\inner_functions.cpp, line 857
Exception in thread "main" CvException [org.opencv.core.CvException: ..\..\..\src\opencv\modules\ml\src\inner_functions.cpp:857: error: (-5) train data must be floating-point matrix in function cvCheckTrainData
]
at org.opencv.ml.CvSVM.CvSVM_1(Native Method)
at org.opencv.ml.CvSVM.<init>(CvSVM.java:80)
I can't manage to train the SVM, any idea? Thanks
Assuming that you know what you are doing by reshaping an image and using it to train SVM, the most probable cause of this is that your
Mat img = Highgui.imread(file.getAbsolutePath());
fails to actually read an image, generating a matrix img with null data property, which will eventually trigger the following in the OpenCV code:
// check parameter types and sizes
if( !CV_IS_MAT(train_data) || CV_MAT_TYPE(train_data->type) != CV_32FC1 )
CV_ERROR( CV_StsBadArg, "train data must be floating-point matrix" );
Basically train_data fails the first condition (being a valid matrix) rather than failing the second condition (being of type CV_32FC1).
In addition, even though reshape works on the *this object, it acts like a filter and its effect is not permanent. If it's used in a single statement without immediately being used or assigned to another variable it will be useless. Change the following lines in your code:
img.reshape(1, 1);
trainingImages.push_back(img);
to:
trainingImages.push_back(img.reshape(1, 1));
Just as the error says, You need to change type of Your matrix, from integer type, probably CV_8U, to floating point one, CV_32F or CV_64F. To do it You can use cv::Mat::convertTo(). Here is a bit about depths and types of matrices.
Related
I came across an error during execute stereoCalibrate in Opencv 2.4.11, which is says :
OpenCV Error: Assertion failed (!fixedSize() || ((Mat*)obj)->size.operator()() == Size(cols, rows)) in cv::_OutputArray::create,
I think this must be some size error between these parameters, which go through them one by one. But there is still error. I hope someone awesome could find the error from the assembly code below. Here is the method call in my code.
double error = cv::stereoCalibrate(
objPoints, cali0.imgPoints, cali1.imgPoints,
camera0.intr.cameraMatrix, camera0.intr.distCoeffs,
camera1.intr.cameraMatrix, camera1.intr.distCoeffs,
cv::Size(1920,1080), m.rvec, m.tvec, m.evec, m.fvec,
cv::TermCriteria(CV_TERMCRIT_ITER + CV_TERMCRIT_EPS, 100, 1e-5)
,CV_CALIB_FIX_INTRINSIC + CV_CALIB_USE_INTRINSIC_GUESS
);
In my code, m.rvec is (3,3,CV_64F), m.tvec is (3,1,CV_64F), m.evec and m.fvec are not preallocated which is same with the stereoCalibrate example. And intr.cameraMatrix is (3,3,CV_64F) and intr.distCoeffs is (8,1,CV_64F), objPoints is computed from the checkerboard which stores the 3d position of corners and all z value for point is zero.
After reading advice from #Josh, I modify the code as plain output mat object which are in CV_64F, but it still throws this assertion.
cv::Mat R, t, e, f;
double error = cv::stereoCalibrate(
objPoints, cali0.imgPoints, cali1.imgPoints,
camera0.intr.cameraMatrix, camera0.intr.distCoeffs,
camera1.intr.cameraMatrix, camera1.intr.distCoeffs,
cali0.imgSize, R, t, e, f,
cv::TermCriteria(CV_TERMCRIT_ITER + CV_TERMCRIT_EPS, 100, 1e-5));
Finally I solved this problem, as a reminder, make sure the camera parameters you passed in are not const type....
Why go for assembly? OpenCV is open source and you can check the code you're calling here: https://github.com/opencv/opencv/blob/master/modules/calib3d/src/calibration.cpp#L3523
If you get assertion fails in OpenCV it's usually because you've passed a matrix with an incorrect shape. OpenCV is extremely picky. The assertion fail is on an OutputArray, so checking the function signature there are four possible culprits:
OutputArray _Rmat, OutputArray _Tmat, OutputArray _Emat, OutputArray _Fmat
The sizing is done inside cv::stereoCalibrate here:
https://github.com/opencv/opencv/blob/master/modules/calib3d/src/calibration.cpp#L3550
_Rmat.create(3, 3, rtype);
_Tmat.create(3, 1, rtype);
<-- snipped -->
if( _Emat.needed() )
{
_Emat.create(3, 3, rtype);
p_matE = &(c_matE = _Emat.getMat());
}
if( _Fmat.needed() )
{
_Fmat.create(3, 3, rtype);
p_matF = &(c_matF = _Fmat.getMat());
}
The assertion is being triggered in one of these calls, the code is here:
https://github.com/opencv/opencv/blob/master/modules/core/src/matrix.cpp#L2241
Try passing in plain Mat objects without preallocating their shape.
I am aware there are several ways to read and write a pixel value of an OpenCV cv::Mat image/matrix.
A common one is the .at<typename T>(int, int) method http://opencv.itseez.com/2.4/modules/core/doc/basic_structures.html#mat-at .
However, this requires the typename to be known, for instance .at<double>.
The same thing applies to more direct pointer access OpenCV get pixel channel value from Mat image .
How can I read a pixel value without knowing its type? For instance, it would be ok to receive a more generic CvScalar value in return. Efficiency is not an issue, as I would like to read rather small matrices.
Kind of. You can construct cv::Mat_ and provide explicit type for elements, after that you don't have to write element type each time. Quoting opencv2/core/mat.hpp
While Mat is sufficient in most cases, Mat_ can be more convenient if you use a lot of element
access operations and if you know matrix type at the compilation time. Note that
Mat::at(int y,int x) and Mat_::operator()(int y,int x) do absolutely the same
and run at the same speed, but the latter is certainly shorter.
Mat_ and Mat are very similar. Again quote from mat.hpp:
The class Mat_<_Tp> is a thin template wrapper on top of the Mat class. It does not have any
extra data fields. Nor this class nor Mat has any virtual methods. Thus, references or pointers to
these two classes can be freely but carefully converted one to another.
You can use it like this
Mat_<Vec3b> dummy(3,3);
dummy(1, 2)[0] = 10;
dummy(1, 2)[1] = 20;
dummy(1, 2)[2] = 30;
cout << dummy(1, 2) << endl;
Why I said 'kind of' in the first place? Because if you want to pass this Mat_ somewhere - you have to specify it's type. Like this:
void test(Mat_<Vec3b>& arr) {
arr(1, 2)[0] = 10;
arr(1, 2)[1] = 20;
arr(1, 2)[2] = 30;
cout << arr(1, 2) << endl;
}
...
Mat_<Vec3b> dummy(3,3);
test(dummy);
Technically, you are not specifying your type during a pixel read, but actually you still have to know it and cast the Mat to the appropriate type beforehand.
I guess you can find a way around this using some low-level hacks (for example make a method that reads Mat's type, calculates element size and stride, and then accesses raw data using pointer arithmetic and casting...). But I don't know any 'clean' way to do this using OpenCV's functionality.
If you already know the type, you can use Mat_<> type for easy access. If you don't know the type, you can:
convert the data to double, so data won't be truncated in any case
switch over the number of channels to access correctly the double matrix. Note that you can have at most of 4 channels, since Scalar has at most 4 elements.
The following code will convert only the selected element of the source matrix to a double value (with N channels).
You get a Scalar containing the value at position row, col in the source matrix.
#include <opencv2/opencv.hpp>
#include <iostream>
using namespace std;
using namespace cv;
Scalar valueAt(const Mat& src, int row, int col)
{
Mat dst;;
src(Rect(col, row, 1, 1)).convertTo(dst, CV_64F);
switch (dst.channels())
{
case 1: return dst.at<double>(0);
case 2: return dst.at<Vec2d>(0);
case 3: return dst.at<Vec3d>(0);
case 4: return dst.at<Vec4d>(0);
}
return Scalar();
}
int main()
{
Mat m(3, 3, CV_32FC3); // You can use any type here
randu(m, Scalar(0, 0, 0, 0), Scalar(256, 256, 256, 256));
Scalar val = valueAt(m, 1, 2);
cout << val << endl;
return 0;
}
I'm new to OpenCV and C++ and I'm trying to build a classifier using Gaussian Mixture Model within the OpenCV. I figured out how it works and got it worked ... maybe. I got something like this now:
If I classify the training samples just after the model was trained and saved, I got the result I want. But when I reclassify my training data using the read(), one of the clusters is missing, means I got different cluster result from the same GMM model. I don't get it now because the cluster I want was gone, I can't reproduce the classification again until I retrained the model using the same data. I checked the code in runtime and the result valule in the Vec2d from which predict() returned was never assigned to 1 (I set 3 clusters).
Maybe there's a bug or I did something wrong?
p.s. I'm using 2.4.8 in VS2013
My programs like this:
train part
void GaussianMixtureModel::buildGMM(InputArray _src){
//use source to train GMM and save the model
Mat samples, input = _src.getMat();
createSamples(input, samples);
bool status = em_model.train(samples);
saveModel();
}
save/load the model
FileStorage fs(filename, FileStorage::READ);
if (fs.isOpened()) // if we have file with parameters, read them
{
const FileNode& fn = fs["StatModel.EM"];
em_model.read(fn);
fs.release();
}
FileStorage fs_save(filename, FileStorage::WRITE);
if (fs_save.isOpened()) // if we have file with parameters, read them
{
em_model.write(fs_save);
fs_save.release();
}
predict part
vector<Mat> GaussianMixtureModel::classify(Mat input){
/// samples is a matrix of channels x N elements, each row is a set of feature
Mat samples;
createSamples(input, samples);
for (int k = 0; k < clusterN; k++){
masks[k] = Mat::zeros(input.size(), CV_8UC1);
}
int idx = 0;
for (int i = 0; i < input.rows; i++){
for (int j = 0; j < input.cols; j++){
//process the predicted probability
Mat probs(1, clusterN, CV_64FC1);
Vec2d response = em_model.predict(samples.row(idx++), probs);
int result = cvRound(response[1]);
for (int k = 0; k < clusterN; k++){
if (result == k){
// change to the k-th class's picture
masks[k].at<uchar>(i, j) = 255;
}
...
// something else
}
}
}
}
I suppose my answer will be too late but as I have encountered the same problem the solution I found may be useful for others.
By analysing the source code, I notice that in the case of EM::COV_MAT_DIAGONAL the eigen values of covariances matrix(covsEigenValues in source code) are obtained via SVD after loading the saved data.
However, SVD computes the singular(eigen in our case)values and stores it in ASCENDING order.
To prevent this , I simply extract directly the diagonal element of loaded covariance matrix in covsEigenValues to keep the good order.
I have a problem over here with some simple stitching tool test using OpenCV.
Here s my code:
IplImage *pLeft,
*pRight;
pLeft = cvLoadImage( "left.jpg" );
pRight = cvLoadImage( "right.jpg" );
cv::Mat cvMatLeft( pLeft, true ),
cvMatRight( pRight, true );
std::vector<cv::Mat> imgs;
imgs.push_back( cvMatLeft );
imgs.push_back( cvMatRight );
cv::Mat cvMatOutput;
cv::Stitcher myStitcher = cv::Stitcher::createDefault( true );
cv::Stitcher::Status myStatus = myStitcher.stitch( imgs, cvMatOutput );
I get back the enum ERR_NEED_MORE_IMGS while running this code.
When i debug into the functions called by OpenCV i did recognize the following uncertainty:
stitch( )'s first argument is an cv::InputArray named images. Taking a closer look at it shows, that the arguments sz.width and sz.height are 0.
Further on running through estimateTransform( ) twice the function matchImages( ) is called where the member imgs_ is checked. This one is derived from the InputArray and has (resulting) the size( ) (of images) being 0.
This leads to the mentioned enum.
What am i doing wrong? Something on initialization of the stitcher or the cv::Mat?
Thanks in advance
I think it occurs when you use similar image. When you use the images which the number of extraced feature points is small, so it does.
So, I'm trying to extract some SURF keypoints, but I want to impose these key points! So, I put the last parameter to "true" which is "useProvidedKeypoints".
Also, when I create my Keypoint, I used the default constructor (so some default values there). I only change the point "pt" and the octave that I set to 3.
I'm using the C++ interface with SURF. But I know that the problem is right at cvExtractSURF because I copied that part of the code in mine to help me debug.
When I call that function, with the last parameter set to true, I got this error:
OpenCV Error: Bad argument (Unknown array type) in cvarrToMat, file /home/widgg/opencv/trunk/modules/core/src/matrix.cpp, line 651
terminate called after throwing an instance of 'cv::Exception'
what(): /home/widgg/opencv/trunk/modules/core/src/matrix.cpp:651: error: (-5) Unknown array type in function cvarrToMat
I really don't know what I'm doing wrong!
EDIT:
Here's some code. First how I create the keypoints (I left a couple of informations, like the layer_id stuff, but you get the main idea):
for (json_pt_info_vector::iterator b_beg = beg->points.begin(); b_beg != b_end; ++b_beg)
{
int layer_id = b_beg->layer_id;
json_point_info_coord &jpic = b_beg->coord;
jpic.feature_id = features[layer_id].keypoints.size();
KeyPoint kp;
kp.octave = 3;
kp.pt.x = jpic.x;
kp.pt.y = jpic.y;
features[layer_id].keypoints.push_back(kp);
}
Here's the call to SURF:
SURF surf(300, 3, 4);
for (int i = 0; i < nb_img; ++i)
{
debug_msg("extract_features #4.1");
cv::detail::ImageFeatures &cdif = features[i];
Mat gray_image = imread(param.layer_images[i], 0); // 0 = force to gray scale!
debug_msg("extract_features #4.2");
vector<float> descriptors;
debug_msg("extract_features #4.3");
surf(gray_image, Mat(), cdif.keypoints, descriptors, true); // MUST BE TRUE TO FORCE THE PROVIDED KEYPOINTS
debug_msg("extract_features #4.4");
cdif.descriptors = Mat(descriptors, true).reshape(1, (int)cdif.keypoints.size());
debug_msg("extract_features #4.5");
gray_image.release();
debug_msg("extract_features #4.6");
images[i] = imread(param.layer_images[i]); // keep the image open
}
It crashes after #4.3 in the debug message!
Hope that helps!
EDIT 2:
I replaced some part by cv::SurfDescriptorExtracter. I replace everything from 4.3 to 4.5 with the following line:
extractor.compute(gray_image, cdif.keypoints, cdif.descriptors);
So now, there's still a bug, but it's located somewhere else, not necessary related to this question!
I'm surprised that the call to surf(gray_image, Mat(), cdif.keypoints, descriptors, true) even compiles. the descriptors argument should be a cv::Mat, not a vector.