##### System information (version)
OpenCV => 4.1.0
Operating System / Platform => Windows10 64 Bit
Compiler => Visual Studio 2017
OpenVINO => w_openvino_toolkit_p_2018.5.456

##### Detailed description
I have trained a mask rcnn model with tensorflow and I used openvino to accelerate it. But I found the
result with using parameter DNN_BACKEND_OPENCV is more accurate than using DNN_BACKEND_INFERENCE_ENGINE.

##### Steps to reproduce

1. Test code is OpenCV_Mask_RCNN.cpp, test images are 2.png and 7.png. All files are packed as files.zip.
   files.zip
2. Model files and config files are here. https://drive.google.com/open?id=1xTI6_bqbZpi3vbeVcFxX1emBSa7QI58z
3. Here is my test result.
   DNN_BACKEND_INFERENCE_ENGINE result
   

DNN_BACKEND_OPENCV result

4. I also test faster rcnn model，both of DNN_BACKEND_OPENCV and DNN_BACKEND_INFERENCE_ENGINE result are same. With mask rcnn model seems bounding rects are same but masks are different. @dkurt This looks like a real bug.