Convertion of custom network from mo (.xml & .bin) to OAK-D not working

NielsOhlsen

Hi all, I have trained a custom CNN with TensorFlow that I can successfully translate with Model Optimizer into .xml and .bin files, and successfully infer with the Intel Inference Engine, inference_request.infer() with good accuracy (position and class).

However, I cannot successfully deploy to OAK-D via the Blob converter (http://blobconverter.luxonis.com/).
I get no error messages, but the OAK-D just outputs a constant prediction.
Has anybody else experienced similar issues ?

erik

Hello NielsOhlsen ,
Does the model expect INT8 or FP16? You have to specify the input layer with -ip when compiling the IR into .blob. Could you also share the script you use to run the inference with this model?
Thanks, Erik

NielsOhlsen

erik Hi Erik, thanks for the answer. I am in contact with Matija, and I have now simplified my network from a 3 tower model to just one tower, taken out a batch normalization layer and that simpler model runs inference on the OAK-D.

So I believe my data formats are correct. I do, BTW use -ip U8 (as default proposed by blob converter).

I will now step by step rebuild my model from something simpler, until I again have issues, then I will get back to Matija, who has proposed we contact Intel Premium Support, as it does seem that Blob converter is not translating my topology correctly.

erik

Hello NielsOhlsen , oh that's interesting, we did see a few issues with openvino (especially with quantization), so this could very well be an issue on their side as well.
Thanks, Erik