Converting model to IR ruins the accuracy

Question

I have a tensorflow LPRnet model that came from this repo here I tested the model as is and after converting it with the model optimizer the frozen model detects the license plates correctly really accurately, while the openvino IR doesn't detect the correct license plate almost ever

any idea as to why the optimization might ruin the model accuracy

Answer 1

The solution to this problem with this specific model was to use the exporter in the repo instead of the optimizer that comes with openvino