It seems a bug in SVD (Bug #1448)

Added by Anatoly Baksheev over 13 years ago. Updated almost 13 years ago.

Status:Done Start date:
Priority:High Due date:
Assignee:Vadim Pisarevsky % Done:

0%
Category:core
Target version:2.4.0
Affected version: Operating System:
Difficulty: HW Platform:
Pull request:

Description

For floats SVD returns wron results (compared with Eigen).

cv::Mat cvA(6, 6, CV_64F, A_data);
cv::Mat cvB(6, 1, CV_64F, b_data);
cv::Mat cvA_float(6, 6, CV_32F, A_data_float);
cv::Mat cvB_float (6, 1, CV_32F, b_data_float);

cv::Mat cvr1, cvr2;
cv::solve(cvA, cvB, cvr1, DECOMP_SVD);
cv::solve(cvA_float, cvB_float, cvr2, DECOMP_SVD);

  • cvr2 !=cvr1*

But in the same time:

Eigen::Map< Matrix<double, 6, 6> > A(A_data);
Eigen::Map< Matrix<double, 6, 1> > b(b_data);

Eigen::Map< Matrix<float, 6, 6> > A(A_data_float);
Eigen::Map< Matrix<float, 6, 1> > b(b_data_float);

Eigen::Matrix<float, 6, 1> res1 = A_float.jacobiSvd(ComputeThinU | ComputeThinV).solve(b_float);
Eigen::Matrix<double, 6, 1> res2 = A.jacobiSvd(ComputeThinU | ComputeThinV).solve(b);

res1 == res2

repro.xml - Test dataset ( 4 matrixes Adata, Bdata, Adata_float, Bdata_float) (2.5 kB) Anatoly Baksheev, 2011-10-26 01:01 pm

testSVD.cpp - C++ file (1.3 kB) João cabral, 2012-06-07 12:52 am

svdtest.m - Matlab/Octave file (254 Bytes) João cabral, 2012-06-07 12:52 am

Related issues

related to Bug #2027: SVD Bug Continue Done 2012-06-07

Associated revisions

Revision 5353b976
Added by Vadim Pisarevsky over 13 years ago

fixed single-precision SVD accuracy on some very ill-conditioned matrices (ticket #1448)

Revision 9a862452
Added by Vadim Pisarevsky almost 12 years ago

added test for bug #1448 and hopefully fixes the bug #2898

Revision 8a962d34
Added by Andrey Kamaev almost 12 years ago

Merge pull request #676 from vpisarev/svd_stuff

added test for bug #1448 and hopefully fixes the bug #2898

Revision 94966b38
Added by Alexander Smorkalov over 11 years ago

Merge pull request #1448 from StevenPuttemans:bugfix_3264

History

#1
Updated by Vadim Pisarevsky over 13 years ago

thanks for the test case! the problem seems to be fixed in r7039. BTW, since A in the test case is symmetrical, for such problems it is recommended to use DECOMP_EIG method, because SVD explicitly processes (A'*A) matrix, which is very, very ill-conditioned in your case.

  • Status changed from Open to Done
  • (deleted custom field) set to fixed

#2
Updated by Andrey Kamaev almost 13 years ago

  • Target version set to 2.4.0
  • Category set to core

#3
Updated by João cabral almost 13 years ago

It seems that this is not totally resolved.
See the file in the attach.

If the matrix is CV_32FC1 sometimes the SVD results are incorrect. In this example the matrix U has some values with the wrong sign, and also it has NaN and inf in some elements. The matrix Vt also has wrong sign in some elements.

If we change all the matrix to CV_64FC1 then the results are accurate.

This happens with the most recent build (booth branch 2.4.1 and trunk 2.4.9 )
Os: Ubuntu 11.10

  • File testSVD.cpp added
  • File svdtest.m added

Also available in: Atom PDF