structured edge detection segfaults on certain images

[raghuramos1987]

When used with images of size 400x225 (wxh) there is a segfault in gradientHist. It looks like heights of 225, 213, 201 fail.

[Dikay900]

please provide an image and the code to reproduce the problem.

[raghuramos1987]

https://gist.github.com/raghuramos1987/ae784241d8b2f6a9483c is the driver program. It needs the image and the model file from structured edge detection.

[raghuramos1987]

This looks like an issue with cvRound in gradientHist where the histogram matrix is created. On some numbers it gets rounded down which ends up writing parts into memory blocks which shouldnt be accessed.

histogram.create( cv::Size( cvRound(src.size().width/float(pSize)),
cvRound(src.size().height/float(pSize)) ),

if src has odd numbered sizes this round can fail to allocate the required size. It should ideally be cvCeil but I dont know if that has any repurcussions on the algorithm itself. (From what I understood in the paper it shouldnt). I can put up a pull request if the original author confirms that this is the case.

Also, all the same test images pass with cvCeil

[mshabunin]

@raghuramos1987, can you, please, provide your cmake output (the part with platform, compiler, configuration)?

Have you tried to reproduce the issue with cvRound only?

const int width = <?>;
const int height = <?>;
const int sz = <?>;
cout << cvRound(width/float(sz)) << ", " << cvRound(height/float(sz)) << endl;

[raghuramos1987]

Yup

This is the result

const int width = 241;                                                                            
const int height = 400;                                                                           
const int sz = 2;                                                                                 
std::cout << cvRound(width/float(sz)) << ", " << cvRound(height/float(sz)) << std::endl;

120, 200

[mshabunin]

@raghuramos1987, cvRound can not produce same results on all platforms, one should not rely on it's output for such tasks. I'd recommend to use cvCeil in this case to avoid invalid memory access.

Feel free to create pull request if this fix will help you.

[berak]

made a standalone test case here: https://gist.github.com/berak/9a556537dccea2709712

[mshabunin]

@berak, thank you for example. As I understand, changing to cvCeil helps, doesn't it? I can make pull request if it fixes the problem (the issue is still can not be reproduced on my machine).

[berak]

yes it does. made #274 already

[mshabunin]

Considering it is done