Bug in inpaint_fsr function for 16 bits / 32 bits / 64 bits images

[ferreram]

System information (version)

• OpenCV => 4.5
• Operating System / Platform => Ubuntu 18.04
• Compiler => GCC

Detailed description

When calling inpaint function from xphoto module with INPAINT_FSR_(BEST/FAST) parameter, there is an OpenCV exception if the provided src image is different than CV_8U.

This behavior is due to the call to : "src.convertTo(src, CV_8U, 1/257.0)" whereas src is a const reference argument.

Steps to reproduce

#include <opencv2/xphoto/inpainting.hpp>

int main()
{
    cv::Mat input_16U_img = cv::Mat::zeros(cv::Size(500,500), CV_16U);
    cv::Mat mask = cv::Mat::zeros(input_16U_img.size(), CV_8U);
    cv::Mat out_img;
    cv::xphoto::inpaint(input_16U_img, mask, out_img, cv::xphoto::INPAINT_FSR_FAST);

    return 0;
}

[alalek]

Please provide complete minimal reproducer as requested by issue's template.

[berak]

imo, this issue is still valid (and should be reopened).
here

opencv_contrib/modules/xphoto/src/inpainting_fsr.impl.hpp

Line 770 in d85887c

src.convertTo(src, CV_8U, 1/257.0);

and here

opencv_contrib/modules/xphoto/src/inpainting_fsr.impl.hpp

Line 785 in d85887c

src.convertTo(src, CV_8U, 255.0);

src is const, but the code tries to convert it in-place, resulting in:

OpenCV: terminate handler is called! The last OpenCV error is:
OpenCV(4.6.0-dev) Error: Assertion failed (!fixedType() || ((Mat*)obj)->type() == mtype) in create, file C:\p\opencv\modules\core\src\matrix_wrap.cpp, line 1166

this issue also surfaced on the forum

and btw, does any of the current buildbots run contrib tests ? (i could not find any)

in case you need a minimal reproducing code:

void f(const Mat &src) {
    src.convertTo(src,CV_8U,255);
}
int main(int argc, char **argv) {
    Mat im(100,100,CV_32FC3,0.2f);
    f(im);
}

(or call xphoto::inpaint() with 16u/32f/64f input)

[3jbrown]

I agree this issue should be re-opened! I can reproduce in c++ or python. INPAINT_FSR_BEST/FAST only work with uint8, nothing larger.