8-point fundamental matrix estimation seems to be broken (Bug #3441)

Added by Alexander Mordvintsev about 11 years ago. Updated almost 10 years ago.

Status:Done Start date:2013-12-18
Priority:High Due date:
Assignee:Vadim Pisarevsky % Done:

100%
Category:calibration, 3d
Target version:3.0
Affected version:branch 'master' (3.0-dev) Operating System:Windows
Difficulty: HW Platform:x64
Pull request:

Description

The attached test_rectify.py script generates a synthetic set of image matches and tries to recover the fundamental matrix with "findFundamentalMat". FM_RANSAC method succeeds, but FM_8POINT fails to give correct answer (which leads to crash on rectification). Data is synthetic and shouldn't contain any outliers.

test_rectify.py (1.9 kB) Alexander Mordvintsev, 2013-12-18 01:20 pm

Related issues

duplicated by Bug #4072: Fundamental Matrix Calculation Done 2014-12-17
duplicated by Bug #4173: 8 point algorithm bug Done 2015-02-05
duplicated by Bug #4186: findFundamentalMat is returning wrong results with FM_8PO... Done 2015-02-09

Associated revisions

Revision eee210f3
Added by Deanna Hood almost 10 years ago

Fix Bug #3441, #4072, #4173: 8-point fundamental matrix calculation error

History

#1
Updated by Anna Kogan about 11 years ago

Hello Alexander,

Thank you for reporting the problem and especially for providing the reproducer for it. If you could work on a solution for the issue on your side, a contribution (see How_to_contribute) would be very appreciated!

#2
Updated by Andrew Senin about 11 years ago

  • Assignee changed from Vadim Pisarevsky to Alexander Mordvintsev
  • Status changed from New to Open

#3
Updated by Vadim Pisarevsky about 10 years ago

  • Assignee changed from Alexander Mordvintsev to Vadim Pisarevsky

#4
Updated by Vadim Pisarevsky about 10 years ago

  • Priority changed from Normal to High
  • Target version set to 3.0

#5
Updated by Jon Waltman about 10 years ago

I would like this fixed too, please! It looks like bug reports #4072 and #4173 suggest the same fix (posted by the same guy).

#6
Updated by Jon Waltman almost 10 years ago

I have verified that the fix mentioned above (and copied below) corrects the issue. For reference, Can Cagman said in bug report #4173:

fundam.cpp - run8Point()
line: 576
gemm( U, Mat::diag(W), 1., 0, 0., TF, GEMM_1_T );
should be changed to
gemm( U, Mat::diag(W), 1., 0, 0., TF, 0);

This is because the SVD call in OpenCV 3.0b fundam.cpp:
SVDecomp( F0, W, U, V, SVD::MODIFY_A );
returns a U matrix that is the transpose of that in OpenCV 2.4.9:
cvSVD( &F0, &W, &U, &V, CV_SVD_MODIFY_A + CV_SVD_U_T + CV_SVD_V_T );
This appears to be the case due to a different set of flags input to the SVD call.

Anyways, the quoted fix above corrects the issue and makes OpenCV 3.0b return the same (correct) results for the 8-point algorithm as OpenCV 2.4.9.

#7
Updated by Deanna Hood almost 10 years ago

  • Status changed from Open to Done
  • % Done changed from 0 to 100

Also available in: Atom PDF