@alalek, I don't think there is easy way to fix the problem. One possible solution would be to store Mat itself inside Input/OutputArray, not the pointer to it, but it would noticeably increase the overhead and then we would have to also store UMat, std::vector etc., which would require a type-safe union. I think, we should just document this behaviour. In other words, I do not think that OpenCV can be modified to make the proposed regression test succeed

@vpisarev Agreed.
I will think about problem detection instead (before current problem with Canny, there were similar problems with inplaces in cvtColor). Probably for debug mode at least (ABI will be changed).

So, the options now are: 1) close this PR, 2) disable the tests for now and merge the PR. What do you prefer?