OpenCV Mat::convertTo(type) does not convert the type

Question

GIVEN:

The following code fragment:

#include <opencv2/core.hpp>
#include <iostream>

int
main(int argc, char** argv)
{

    cv::Mat a = (cv::Mat_<double>(3,1) << 1,   2,   3); 
    cv::Mat b;

    std::cout << "a(before): " << cv::typeToString(a.type()) << std::endl;
    std::cout << "b(before): " << cv::typeToString(b.type()) << std::endl;

    std::cout <<                                             std::endl;
    std::cout << "Convert 'a' --> 'b' with type CV_64FC4" << std::endl;
    std::cout <<                                             std::endl;
    a.convertTo(b, CV_64FC4);

    std::cout << "a(after):  " << cv::typeToString(a.type()) << std::endl;
    std::cout << "b(after):  " << cv::typeToString(b.type()) << std::endl;

    return 0;
}

EXPECTATION:

Should produce, in my understanding, the following output:

a(before): CV_64FC1
b(before): CV_8UC1

Convert 'a' --> 'b' with type CV_64FC4

a(after):  CV_64FC1
b(after):  CV_64FC4

OUTPUT:

Instead, the output is as follows:

a(before): CV_64FC1
b(before): CV_8UC1

Convert 'a' --> 'b' with type CV_64FC4

a(after):  CV_64FC1
b(after):  CV_64FC1

QUESTION:

What is going on here? How can I actually convert to the specified target type?

wohlstad · Accepted Answer · 2022-06-23T13:05:32.393

2

Short answer:
cv::Mat::convertTo does not support changing the number of channels.

Longer answer:
As you can see in the documentation regarding the rtype paremeter of cv::Mat::convertTo (the one you pass CV_64FC4 to):

desired output matrix type or, rather, the depth since the number of channels are the same as the input has; if rtype is negative, the output matrix will have the same type as the input.

I.e. convertTo does not handle the case of changing the number of channels (only the bit depth of each channel).
Although it is not documented explicitly, I guess cv::Mat::convertTo extracts the bit depth from rtype and ignores the number of channels.

In your example: a has a single channels, and therefore so is b after the conversion.
In order to see the effect of convertTo, you can pass e.g. CV_32FC1 in order to convert 64 bit doubles to 32 bit floats.

Update:
According to the OP's request (in the comments below), here are examples of changing the number of channels using cv::mixChannels:

cv::Mat img1c = (cv::Mat_<double>(3, 1) << 1, 2, 3);
cv::Mat img4c(img1c.size(), CV_64FC4);
cv::Mat img1c_2(img1c.size(), CV_64FC1);

// Convert 1 channel into 4 (duplicate the channel):
std::vector<int> fromTo1{ 0,0, 0,1, 0,2, 0,3 }; // 0->0, 0->1, 0->2, 0->3
cv::mixChannels(img1c, img4c, fromTo1);

// Convert back to 1 channel (pick one of the channels - the 1st one in this case):
std::vector<int> fromTo2{ 0,0 };    // 0->0  (pick the 1st channel)
cv::mixChannels(img4c, img1c_2, fromTo2);

edited Jun 23 '22 at 13:05

answered Jun 23 '22 at 11:53

wohlstad

12,661
10
26
39

Hm, then why does `convertPointsToHomogeneous(vectorPoint3d, x)` produce a matrix `m` with four channels? Seriously, I am having trouble multiplying `x` here with other matrices of type `CV_64FC1`. – Frank-Rene Schäfer Jun 23 '22 at 12:00
I haven't used `convertPointsToHomogeneous`, but it is a different method. Why do you think it says something about the behavior of `convertTo` ? – wohlstad Jun 23 '22 at 12:02
The thing is, that a matrix (N, 4) has already four homogenous coordinates per row indicating a point. Having four channels does not make sense. Why would that thing create three more channels? How could I possibly rid them? – Frank-Rene Schäfer Jun 23 '22 at 12:06
1

Not sure I understood. You can use `cv::mixChannels` to convert a 1 channel into 4 (or the other way around). If you are not familiar with it I can give an example. – wohlstad Jun 23 '22 at 12:10
yes, please give me an example. – Frank-Rene Schäfer Jun 23 '22 at 12:13
1

@Frank-ReneSchäfer - see examples in my updated answer. – wohlstad Jun 23 '22 at 12:18
1

mixChannels is unnecessary or might not even do what he needs. there is `reshape()` and there is `convertPointsFromHomogeneous()`, as I've explained in the [other question](https://stackoverflow.com/questions/72727013/opencvmat-how-to-efficiently-add-a-column-of-ones) – Christoph Rackwitz Jun 23 '22 at 13:05
1

@Frank-ReneSchäfer you need to learn about the difference between a `cv:Mat` of shape Nx1 with ***K-channel*** elements, and a `cv::Mat` of shape ***NxK*** with 1-channel elements. do you see that they are nearly equivalent? OpenCV treats those as equivalent in most situations, and prefers the former. – Christoph Rackwitz Jun 23 '22 at 13:11

Christoph Rackwitz · Answer 2 · 2022-06-23T13:13:08.033

As I've commented in your other question, if you want to rearrange a Nx1 4-channel Mat to Nx4 1-channel, then you need to use cv::Mat::reshape()

some_mat.reshape(1) converts from Nx1 K-channel (CV_64FC4) to NxK 1-channel (CV_64FC1)

Docs: https://docs.opencv.org/4.x/d3/d63/classcv_1_1Mat.html#a4eb96e3251417fa88b78e2abd6cfd7d8

You really don't need any of that though, because you say you want matrix multiplication... or rather matrix-vector products. you can have that with cv::transform. transform does matrix-vector multiplication, also for lists of vectors (Nx1 K-channel matrices, also NxK 1-ch).

OpenCV Mat::convertTo(type) does not convert the type

2 Answers2