Deploy pretrained model on OAK-D device

Hello @Areeba,
that's awesome to hear! We have a tutorial on how to convert trained models into IR/blob for deployment to the DepthAI here. There are tons of useful resources linked from that page that can help you
Thanks, Erik

Hello @Areeba,
So you have to load your blob into the neural network node and later get the result of the NN. This demo does a similar thing - loading a blob and decoding the result.
Thanks, Erik

Hello @Areeba ,
I believe the issue is with the NN result decoding. As the error states, size of the bounding boxes array is 0, and you can't access the first element. Different NN models can have a different output so you have to change the decoding part of the code.
Here are some examples of decoding the NN result, just to give you an idea:

• object detection (efficientDet)
• classification (fire detection)
• semantic segmentation (deeplab)

Thanks, Erik

Hello @Areeba ,
so you used the tiny yolo v3 decoding example, modified parts that are specified by the 3 bullet points and it returns you the XlinkOut 220 error? Or is it resolved already? I tried those steps myself (when writing the instructions) and it worked flawlessly, sorry for the troubles.
Thanks, Erik

Just tried it out, and it throws the XlinkOut 220 error as well - which means the output of the NN is too big. I would say there was some problem when converting the NN - resulting in the output of the blob being too big.
We will have to troubleshoot the notebook, and we actually were planning on updating all notebooks, so we will check for this issue as well. Thank you for reporting it, we will circle back when notebooks are updated
Thanks, Erik

@Areeba we will most likely update them next week. Until then, I recommend using https://roboflow.com/, they even have a tutorial on how to use their platform with depthai. Hope it helps

Areeba the notebooks were updated to use latest OpenVINO and blobconverter tool, feel free to give them a try

Hi...To prepare your possess custom demonstrate, you must assemble a dataset of pictures, clarify them, prepare your show, and after that change over and optimize your show for your arrangement goal. This machine learning pathway is fashioned with subtlety, and moderates advancement cycles. What’s more – botches don't show themselves not as obtrusive mistakes, but or maybe, unobtrusively corrupt your model’s performance. In exertion to streamline computer vision advancement and deployment, Roboflow, OpenCV, and Luxonis are collaborating to supply a one-click custom preparing and sending arrangement to the OAK-1 and the OAK-D.

Areeba did you follow the "Run on the depthai device" section of the notebook (at the bottom)? That should work as expected.