Tests fail on i386

[dkogan]

Hello. The Debian g2o package is failing on i386. The build log: https://buildd.debian.org/status/fetch.php?pkg=g2o&arch=i386&ver=0%7E20230806-2&stamp=1696420212&raw=0

It looks like there's a test failure:

./unit_test/slam3d/optimization_slam3d.cpp:94: Failure
Value of: v2AfterOpti->estimate().translation().isApprox(g2o::Vector3::Zero())
  Actual: false
Expected: true

Can we please improve the error message? I suspect that this is simply an issue of numerical fuzz, and increasing the threshold used by isApprox() would fix this. It would be nice if the error message clearly said what threshold it's using, and what value it saw instead.

Thanks

[RainerKuemmerle]

Agree, we should use test code that better prints actual and expected value.
Will look at this, in particular, because on https://eigen.tuxfamily.org/dox/classEigen_1_1DenseBase.html#ae8443357b808cd393be1b51974213f9c I found the interesting note that it does not work to check for approximately zero.

[RainerKuemmerle]

Pushed a commit with which a failing test would print, e.g.,

/home/goki/workspace/g2o/unit_test/slam3d/optimization_slam3d.cpp:139: Failure
Value of: print_wrap(v2AfterOpti->estimate().rotation().diagonal())
Expected: does differ from { 0, 0, 0 }
more than 0.001
  Actual: { 0.99999999999999989, 0.99999999999999989, 0.99999999999999989 }

[dkogan]

Hi. I cherry-picked that patch into the g2o package, and it builds on i386 now. To confirm: that patch adds extra diagnostics AND loosens the acceptance threshold? Yes? Thanks for patching this. I'm closing the bug.

[RainerKuemmerle]

Yes. With the patch we add diagnostics and an adapted acceptance threshold for the test.