@ivashmak, @dkurt, PTAL

There is actually another bug at HEAD: H is not normalized if it is refined.
We need a H.convertTo(H, H.type(), scaleFor(H.at<double>(2,2))); here:

@vrabaud, I recommend to add all findings to PR, thanks.

Unfortunately, I also cannot find a proper test. Let's review by definition if you have no test case to catch the problem.

@dkurt, done. BTW, shouldn't we normalize H by dividing it by H.at<double>(2,2) instead of multiplying by the inverse ? I understand divisions are still 4/5 times slower according to uops.info but at least, we would get exactly 1 for H(2,2).

@vrabaud, please review commit 3b90aa0 that adds two checks for H normalization and Jacobian calculation for the refinement process.

@dkurt, thx for the extra commit, +1