C++ OpenCV - 创建 3D 矩阵并访问其元素

Question

I'm trying to use a 3D matrix in OpenCV to store and access data (type float). Currently I have 3 methods to create the 3D matrix itself of size 158 x 98 x 32 and initialize it at zero:

int out[3];
out[0] = 98;
out[1] = 158;
out[2] = 32;

//Alternative 1:
cv::Mat M(3, out, CV_32FC1, cv::Scalar(0));

//Alternative 2:
cv::Mat M(3, out, CV_32F);
M = Scalar(0);

//Alternative 3:
Mat *feat = new Mat(3,out,CV_32F,Scalar(0));
Mat M = *feat;

Then I use the .data function to obtain a pointer to the first data element:

unsigned char *input = (unsigned char*)(M.data);

Next, I thought I could set the first element (0,0,0) to 1 by using either of the following two methods:

input[0]= 1;             //Alternative 1
M.at<float>(0,0,0) = 1;  //Alternative 2

The at() method works just fine, but I can't seem to get the pointer to work. The following:

input[0]= 1;
input[1]= 2;

Results in 4.32e-42 at element (0,0,0) while the other command seems to have no effect at all. According to the documentation, input[0] should refer to the value at point (0,0,0), input[1] to (0,0,1) and input[32] to (0,1,0) etc.

Also, the M.step is set to 0 and M.cols and M.rows are both -1. The number of rows and columns seems right for a multidimensional matrix, but the step should have a value, right?

So, which of the 3 alternatives is best to initialize the 3D matrix and how can I use pointers to access each data elements separately?

By the way, I'm using the following code to output the contents of the matrix (any other ideas are welcome as well):

float M_res = 0;
ofstream res;
res.open("Results.txt"); //Open file

for(int loopz=0;loopz<out[2];loopz++) {
  res << endl << endl << "Dimension " << loopz << endl;
    for(int loopy=0;loopy<out[0];loopy++) {
      res << endl;
      for(int loopx=0;loopx<out[1];loopx++) {
        M_res = M.at<float>(loopy,loopx,loopz)
        res << M_res << " ";
      }
    }
}

Answer 1

Hi, I don't have enough reputation to comment and since there is neither answers nor comments I will write my comment here.

It seems like this line is erroneous:

Then I use the .data function to obtain a pointer to the first data element:

unsigned char *input = (unsigned char*)(M.data);

since the matrix M type is float therefore you should use

float *input = (float *)(M.data);

Answer 2

I know this is an old post, but I write the answer to whomever comes and passes by.

Two separate questions: initializing matrices and accessing multidimensional array elements.

---

For the first question , on initializing matrices, I would state that all three are functional. However, the first is the best one since it is encapsulated in a standard cv constructor. The second one I would say that is exactly an homologue, since you are creating a matrix which is an object with a data object inside. Stating M = Scalar(0) is just initializing all values of the data object inside M , as in alternative 1. The third case is like the first, but through a pointer. To provide more clarity on the subject, I hope the next paragraph from Learning OpenCV 3: Computer Vision in C++ with the OpenCV Library :

In addition, some of these [constructors] allow you to initialize the data, either by providing a cv::Scalar (in which case, the entire array will be initialized to that value), or by providing a pointer to an appropriate data block that can be used by the array. In this latter case, you are essentially just creating a header to the existing data (ie, no data is copied; the data member is set to point to the data indicated by the data argument).

---

Regarding the second question , accessing elements. We are dealing with a single valued 3 dimensional matrix. To access the elements individually, there are different options: by location or iteration.

Location supports two standard ways: at<>() and ptr<>() . The second is slightly faster than the first. To put in a example, as you already put an example for the first:

M.ptr<float>(3) will access the address of the first element in row 3. With this and knowing if values are conitnuous in memory (with isContinuous ), you can run through it by accessing the values, *M.ptr<float>(i) .

For example, while refactoring your code:

    int out[3] = {98, 158, 32};

    //Alternative 1:
    cv::Mat M(3, out, CV_32F, cv::Scalar(0));

    cout << "Total size " << M.size <<" and, for example, rows " << M.size[0] << endl;

    cout << "IS the matrix really continuous? " << M.isContinuous() << " Yes" << endl;
    int counter = 0;
    for (int r = 0; r < M.size[0]; r++){
        float* p = M.ptr<float>(r);
        for (int c = 0; c < M.size[1]*M.size[2]; c++){
            *(p + c);
            counter += 1;
        }
    }
    cout << "Number of elements seen " << counter << " matrix number of elements " << 98*158*32 << endl;

Outputs:

Total size 98 x 158 x 32 and, for example, rows 98
IS the matrix really continuous? 1 Yes
Number of elements seen 495488 matrix number of elements 495488

The last way to access elements is by the use of iterators, which work as in STL.

    cv::MatConstIterator_<float> it = M.begin<float>();

    while ( it != M.end<float>()){
        *it;
    }

But accessing by pointer is faster.

Also, to access them in group you can check the same book, chapter 4, Section Accessing Array Elements by Block , or the OpenCV docs for methods of m.row , m.col , and others, where m is cv::Mat .