Thanks for the info, I will look further into local conversion. I have already trained two custom models using the colab notebook provided. They both perform well, but one of them is not quite as good as the SSD MobileNet V2 FPNLite 320*320 that I previously created. As it stands, my training images where taken with a different camera so I am adding images from my OAK-D and will retest. This by itself may be enough to improve the accuracy.