@tsocha @mandrono could you please take a look?
I believe this change is not needed anymore because the NMS-5 is already supported on CPU (#2768). @mandrono am I right?
BTW, the initial change with generating NMS-4 looked strange from my perspective, because onnx_importer is a common part, and used not only by CPU plugin. This change has broken very critical models for VPU plugin because we have already moved to NMS-5 usage.

@AndrewBakalinIntel
Originaly the reason why ONNX Importer switched to NMS-4 is that because NMS-5 was merged into master without required transformations.
As you can see you have to xfail a model when NMS-4 is changed to NMS-5 and this is a regression, we should avoid this IMO.

From the logs I can see that there is a type conversion missing.
Could you add it to this file?
https://github.com/AndrewBakalinIntel/openvino/blob/ab/return-nms-5/inference-engine/src/transformations/src/transformations/convert_precision.cpp#L224-L238

@tsocha @postrational CI is still failing, most probably because #2450 is still not merged.
It is a big problem that the operation was merged without appropriate transformations, but I still think we need to keep generating NMS-5 and solve the problem appears instead of masking them (at least it will help us to track all the current issues), because it anyway should be supported on CPU for the upcoming release. So I suggest to suppress the failing models (as it was actually for some time before), it will unblock a lot of critical models for VPU and we will continue to track the remained issues.

@AndrewBakalinIntel
Yes, NMS-5 is already supported on CPU plug-in