GpuMat::convertTo() is unable to convert CV_8U to CV_16F

C++
1

The result of the conversion is only 0s!

Sample code:

cv::Mat cpuTmp, cpuMat(32,32,CV_8UC3,cv::Scalar(255,0,0)); //b,g,r

cv::cuda::GpuMat gpuMat, gpuMat16;
gpuMat.upload(cpuMat);

gpuMat.convertTo(gpuMat16, CV_16F);
gpuMat16.download(cpuTmp);

The result into cpuTmp contains only zeroes!

Any idea how to solve this, or maybe there is another way for this conversion into the GPU memory?
PS: I should have to mention, that the OpenCV version I use is 4.5.4.

2

I don’t think it works on the CPU either.

3

No, it works fine, I have dumped and checked the result. But CV_16F on GPU produces only 0000000…

4

Apologies, I was examining with Image Watch.

Appears that when that type was introduced the CUDA functions were not updated to deal with it.

5

“Appears that when that type was introduced the CUDA functions were not updated to deal with it.”

That’s right - the type code is 7, and this was formely OTHER, as I understand during googling for a solution.

6

cuda modules live in the contrib repo. you could file a bug there (but check for existing issues first).

opencv type codes are about to experience some overhaul. the element type is in the lower 3 bits so far (CV_MAKETYPE and CV_CN_SHIFT). I saw talk of extending that space. it’ll break some binary compatibility. they might do that for OpenCV v5, or maybe later.