cv::meanStdDev returns inf sd for CV_32FC1 Mat in release build (Bug #4194)

Added by Sergey Missan about 10 years ago. Updated almost 10 years ago.

Status:Cancelled Start date:2015-02-12
Priority:Low Due date:
Assignee:- % Done:

0%
Category:core
Target version:-
Affected version:2.4.9 (latest release) Operating System:Windows
Difficulty: HW Platform:x86
Pull request:

Description

when using cv::meanStdDev on CV_32FC1, the mean is calculated correctly, but sd is always inf. This works correctly in debug build (both mean and sd are calculated) but fails in release build.

Example code

cv::Scalar     mean;
  cv::Scalar     stddev;
  cv::meanStdDev(im, mean, stddev);

stddev.val0 will be inf!

the range of values in the im is 5.0921e+13 4.95419e+20

History

#1
Updated by Nisarg Thakkar about 10 years ago

Hi Sergey,
Can you attach the im that you used?

#2
Updated by be rak about 10 years ago

imho, we should not use 32bit floats to represent values in the [5.0921e+13 4.95419e+20] range, but i'm able to reproduce it:


Mat m(5,5,CV_32F);
randu(m,Scalar(1.0e13),Scalar(1.0e20));

cv::Scalar     mean;
cv::Scalar     stddev;
cv::meanStdDev(m, mean, stddev);

cerr << mean << endl;
cerr << stddev << endl;

debug:

[4.57531e+019, 0, 0, 0]
[2.96023e+019, 0, 0, 0]

release:
[4.57531e+019, 0, 0, 0]
[1.#INF, 0, 0, 0]

#3
Updated by Nisarg Thakkar almost 10 years ago

This seems a problem in IPP. If I remove the IPP for CV_32FC1, the result is calculated correctly. But, if the numbers reduce, IPP seems to work alright.
This code is present in the function cv::meanStdDev in the file modules/core/src/stat.cpp at the line:

#if (IPP_VERSION_X100 >= 801)
type == CV_32FC1 ? (ippiMeanStdDevFuncC1)ippiMean_StdDev_32f_C1R ://Aug 2013: bug in IPP 7.1, 8.0
#endif

#4
Updated by Vadim Pisarevsky almost 10 years ago

  • Category set to core

#5
Updated by Vadim Pisarevsky almost 10 years ago

I agree with be rak that double precision should be used for such a range. working range for single precision operations is ~[-10**38, 10**38]. since meanStdDev sums the squared values as intermediate result, the range [10**13, 10**20] becomes [10**26, 10**40], which is too much. I doubt that it can be fixed. Even if we disable IPP, further optimizations of the C branch may yield the same "bug" again. So it's easier to say now - no, we are not going to handle such cases.

  • Status changed from New to Cancelled
  • Priority changed from Normal to Low

Also available in: Atom PDF