Is Mask RCNN compatible to run in Intel NCS2?

[dilip96]

HI,

I already have a setup of NCS2 with Raspberry Pi with SSD MobileNet running.

I need the Mask RCNN model for my applications. How compatible is it?

[David_C_Intel]

Hi dilip96,

Thanks for reaching out.

Mask RCNN is not supported by the NCS2. As you can check here, in the note it states:

NOTE: On VPU devices (Intel® Movidius™ Neural Compute Stick, Intel® Neural Compute Stick 2, and Intel® Vision Accelerator Design with Intel® Movidius™ VPUs) this demo is not supported with any of the Model Downloader available topologies. Other models may produce unexpected results on these devices as well.

Best regards,

David C.

[dilip96]

Hi David,

Thank you for your response.

Is there any solution that you can provide me for running Mask RCNN model?

I am looking to develop a compact Device With AI capabilities. Mask RCNN seems to be necessary for my application.

Can you suggest me any other hardware accelerators for this case.

Thanks again!!

[dilip96]

Also do we know the reason why The Mask RCNN model is not supported.

Is that the VPUs don't support the computational requirement of MaskRCNN?

If we have a custom trained model of MaskRCNN with MobileNet backbone (Which I feel will reduce the Model size.), is there a chance it could work??

Thanks Again !!

[David_C_Intel]

Hi dilip96,

Regarding your questions:

• Is that the VPUs do not support the computational requirement of MaskRCNN?

That is correct, the NCS does not support Mask RCNNs as the device does not have enough memory to allocate tensors on VPU.

• If we have a custom trained model of MaskRCNN with MobileNet backbone (Which I feel will reduce the Model size.), is there a chance it could work??

Although we cannot guarantee it would work, there is a chance it might work if:

1. All layers on your model are supported by the VPU. You can refer to this Supported Layers documentation for more information.
2. Model it is within the memory limit of the device.

Once you have your model, you can try loading the model to the NCS and see if it times out. If the model uses more memory beyond the memory available on NCS, you would see an error message similar to "ncGraphAllocate:2127 Not enough memory to allocate intermediate tensors on remote device".

Best regards,

David C.

[dilip96]

Hi David,

Thanks for your assistance. Ill get back to you if there are any other queries.

[Valerien]

Hey, im trying to use mask R-CNN on NCS2 also too.Did you succeed to run ? 

[corlangerak]

I am looking for mask-rcnn too! Did you manage to get it working?