They have a converter tool that allows you to convert your tf and tf lite models to the format that can be used by KPU. So, the solution here would be to train your NN for recognizing specific objects you need. I plant to publish a tutorial on how to do that as well