@dkurt Please, take a look at the fix (seems like a crutch, but anyway). Still, it doesn't fix the problem on whole, because there are many unsupported operations in the ONNX model:
Control Flow:

Eltwise:

They cannot be reduced to the current implementations of Detection output layers, so I think it is a good idea to implement these layers as was suggested in #16750.
Can we discuss it somewhere outside Github?

Even all the layers will be imported, model execution will be inefficient because of too atomic operations.

Even all the layers will be imported, model execution will be inefficient because of too atomic operations.

Yes, I understand it, but to prevent it I plan to implement aggregate layers for RCNN models in a way similar to current Softmax or BatchNormalization OpenCV layers. And even in this case extra layers are required.

@mshabunin I think there is a problem with Linux Custom builder, there might be some problems with MYRIAD device, could you check?

@sl-sergei , there is a merge conflict in opencv_extra repository now, please rebase your branch.

Looks like this patch brokes MYRIAD tests somehow:

• good: http://pullrequest.opencv.org/buildbot/builders/3_4_openvino-skl-lin64/builds/857 ( 564d1a0 )
• bad: http://pullrequest.opencv.org/buildbot/builders/3_4_openvino-skl-lin64/builds/856 ( 2b82f8f )