ML Object detection / tracking - how to make a custom .onnx file?

Ines Hilz

Hi all!

I am wildly fascinated by the ML features, and made some object detection experiments with the coco datasets before.

Now clients are asking for specific objects, like statues, or a product packaging.
For the latter i prepared a dataset with these tools:

www.lobe.ai
www.teachablemachine.withgoogle.com
www.edgeimpulse.com

but the formats that i was able to export there didn't match with Lens studio.
With Lobe i can even export a .ONNX or a .PB file, but if i place it in Lens Studio, i get this error messages:

.onnx :
Resource import for … failed: ONNX model Conv layer converter does not support padding mode NOTSET

.pb:
Resource import for … failed: Unable to parse TensorFlow proto.

Did anyone manage to make a custom ML Detection yet? If so, how??
Thanks !

Askhat Mudarisov

Lens Studio converts the model with their own converter, and the error messages are not very useful.

Check https://docs.snap.com/lens-studio/references/guides/lens-features/machine-learning/compatibility/

Supports CONSTANT, REFLECT and SYMMETRIC padding types. Seems like your models have : "NOTSET"