Export Mask R-CNN to ONNX format

I am considering using MATLAB to train a Mask R-CNN network but the trained network must be able to be exported to ONNX format afterwards so I can import and use it in my application.

I took a look at the exportONNXFormat function and exporting a roiAlignLayer doesn't seem to be listed. Can anyone confirm whether this kind of layer is compatible with the ONNX export?

Otherwise I can probably use a Fast or Faster R-CNN for my application, and the roiMaxPooling2D layer is explicitly mentioned in the function documentation as being supported.

Thank you

Answer 1

[Sourabh]

Hey Karl,

As you have mentioned, the roiAlignLayer is not supported by the exportONNXNetwork function currently. I have attached a few relevant links which might help get you started with your approach using a Fast or Faster R-CNN network:

• https://www.mathworks.com/help/releases/R2022a/vision/ug/getting-started-with-r-cnn-fast-r-cnn-and-faster-r-cnn.html
• https://www.mathworks.com/help/releases/R2022a/vision/ug/object-detection-using-deep-learning.html
• https://www.mathworks.com/help/releases/R2022a/vision/ug/object-detection-using-faster-r-cnn-deep-learning.html

All the best!

[Karl Mueller]

Thank you,

I have also found that one of the layers in a faster R-CNN is not supported by the ONNX export.

I have designed a fast R-CNN and that should be ok according to the documentation.