Reputation: 2102
Im having an issue using opencv trying to convert an image to an array. The conversion works however i seem to have incorrect dimensions in the resulting array:
#include <opencv2/opencv.hpp>
int main()
{
auto img = cv::imread("test.jpg", CV_LOAD_IMAGE_COLOR);
std::cout << "img cols: " << img.cols << " img rows: "
<< img.rows << " channels: " << img.channels() << std::endl;
std::vector<float> array2;
if (img.isContinuous()) {
array2.assign((float*)img.ptr(0), (float*)(img.ptr(img.rows - 1)) + img.cols);
std::cout << array2.size() << "\n";
}
return 0;
}
The output from the first print line results in :
img cols: 416 img rows: 416 channels: 3
Which is correct, however after assigning the data to the array the dimensions are : 518336 , when they should be 519168 (416*416*3).
Could anyone possibly suggest what exactly is causing the resulting array to be smaller than expected?
Upvotes: 0
Views: 1382
Reputation: 19071
There are several problems with your code:
First of all, cv::imread("test.jpg", CV_LOAD_IMAGE_COLOR);
will (on success) return a cv::Mat
with datatype CV_8UC3
, however you're accessing the elements as float
s. This means that the values you will read will be garbage, and you will also end up reading past the end of the pixel buffer.
If you want floats, then you need to do some conversion/casting, either before or during the act of copying.
The second problem lies in your calculation of the "end" pointer, where you seem to forget that you're dealing with a multi-channel cv::Mat
. In case of a CV_8UC3
matrix, each pixel is represented by 3 bytes, hence there are cols*channels
bytes per row. (That's why you're short by 2*416 elements)
Not really a problem, but a limitation -- your code only works for continuous Mat
s.
I would take a somewhat different approach, and take advantage of functionality provided by OpenCV.
Use cv::Mat::copyTo
, since OutputArray
can wrap a std::vector<T>
. However, for this to work, the source Mat
needs to have 1 channel and 1 row. We can achieve this efficiently using cv::Mat::reshape
, but the Mat
needs to be continuous, so that limitation stays.
std::vector<uchar> to_array_v1(cv::Mat3b const& img)
{
std::vector<uchar> a;
if (img.isContinuous()) {
img.reshape(1, 1).copyTo(a);
}
return a;
}
Use MatIterator
s which we can get using cv::Mat::begin
and cv::Mat::end
. The iterators will work correctly even on a non-continuous Mat
, however we need them to iterate over bytes, so we need to reshape
the matrix to a single channel one. Since we're not changing the number of rows, the reshape will also work on a non-continuous Mat
.
std::vector<uchar> to_array_v2(cv::Mat3b const& img)
{
cv::Mat1b tmp(img.reshape(1));
return std::vector<uchar>(tmp.begin(), tmp.end());
}
The approach suggested by Silencer, using the rather poorly documented cv::Mat::datastart
and cv::Mat::dataend
members. The documentation of cv::Mat::locateROI
sheds some more light on the meaning of those member variables:
However, each submatrix contains information (represented by
datastart
anddataend
fields) that helps reconstruct the original matrix size and the position of the extracted submatrix within the original matrix.
This means that this approach has 2 limitations: it needs a continous matrix, and it won't work correctly for a submatrix, even if it's continuous. (Specifically, for a continuous submatrix, it would return the entire buffer of the "parent" matrix)
std::vector<uchar> to_array_v3(cv::Mat3b const& img)
{
std::vector<uchar> a;
if (img.isContinuous() && !img.isSubmatrix()) {
a.assign(img.datastart, img.dataend);
}
return a;
}
#include <opencv2/opencv.hpp>
#include <iostream>
#include <numeric>
#include <vector>
// Paste implementations from the answer here
cv::Mat3b test_image()
{
cv::Mat1b m(4, 4);
std::iota(m.begin(), m.end(), 0);
cv::Mat3b img;
cv::merge(std::vector<cv::Mat1b>{ m * 3, m * 3 + 1, m * 3 + 2 }, img);
return img;
}
void print(cv::Mat3b const& img)
{
std::cout << "Continuous: " << (img.isContinuous() ? "yes" : "no") << '\n';
std::cout << "Submatrix: " << (img.isSubmatrix() ? "yes" : "no") << '\n';
std::cout << img << "\n";
}
void print(std::vector<uchar> const& a)
{
if (a.empty()) {
std::cout << "empty";
} else {
for (auto n : a) {
std::cout << int(n) << ' ';
}
}
std::cout << "\n";
}
void test(cv::Mat3b const& img)
{
print(img);
print(to_array_v1(img));
print(to_array_v2(img));
print(to_array_v3(img));
}
int main()
{
cv::Mat3b img(test_image());
test(img);
cv::Mat3b img2(img(cv::Rect(0, 0, 3, 3)));
test(img2);
cv::Mat3b img3(img(cv::Rect(1, 1, 3, 1)));
test(img3);
return 0;
}
Running this program will produce the following output:
Continuous: yes
Submatrix: no
[ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11;
12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23;
24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35;
36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47]
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
Continuous: no
Submatrix: yes
[ 0, 1, 2, 3, 4, 5, 6, 7, 8;
12, 13, 14, 15, 16, 17, 18, 19, 20;
24, 25, 26, 27, 28, 29, 30, 31, 32]
empty
0 1 2 3 4 5 6 7 8 12 13 14 15 16 17 18 19 20 24 25 26 27 28 29 30 31 32
empty
Continuous: yes
Submatrix: yes
[ 15, 16, 17, 18, 19, 20, 21, 22, 23]
15 16 17 18 19 20 21 22 23
15 16 17 18 19 20 21 22 23
empty
Upvotes: 2
Reputation: 18341
Mat img = imread("test.png");
std::vector<uchar> arr;
// convert Mat of CV_8UC3 to std::vector<uchar> if continuous
if(img.isContinuous()){
arr.assign(img.datastart, img.dataend);
}
Upvotes: 0