3D CNN into Movidius NCS

[idata]

Hello everyone!

I'm working with a CNN object recognition network with 3D voxel inputs (30x30x30) like the image.

The model is in TensorFlow. When I am using mvNCCompile to generate .graph file I get the following error:

[Error 5] Toolkit Error: Stage Details Not Supported: Conv3D

Is there any possible solution?

Thanks!

[idata]

@CristianWP At the moment, the Neural Compute SDK doesn't have support for the 3D convolution layer in TensorFlow. There isn't a workaround for this at the moment. You can try using Intel's OpenVINO toolkit and see if this toolkit can work for your network: https://software.intel.com/en-us/openvino-toolkit. Here is a link to the OpenVINO developer forum: https://software.intel.com/en-us/forums/computer-vision.

[idata]

Thank you very much Tome!

[idata]

Hi Tome,

My CPU does not meet the requirements for OPENVINO.

Is there another way to generate Movidius .graph file without using mvNCCompile?I am currently using Movidius NCS (with myriad 2), is there any tool in Myriad Development Kit & Board or in Google Vision AIY Kit that can allow me to implement 3Dconvolutions?

Thanks!

[idata]

@CristianWP OpenVINO lists the CPU requirements as 6-8th gen Intel CPU, but you should give it a try regardless.

There isn't another method of generating a compatible graph file without using the NCSDK mvNCCompile tool or OpenVINO's Model Optimizer compiler. The Google AIY Vision Kit is not compatible with the NCSDK and the MDK is only available to high volume sales customers.