Good work David_sh. I also came across some slowness with Tensorflowsharp when I tried out image recognition, and like yourself I’m not sure what the reason for the slowness was. It seemed to get better after the first image, perhaps because it was the start of the session, or maybe once I sorted out the garbage collection in a later version. That the model runs faster in Python may also be related to the fact the Python API and the C API actually reference different c++ code bases, with the C API codebase as the redheaded step child of the suite.
Also a word of caution:
The Tensorflow C API that Tensorflowsharp is based on is incomplete with some functions available but not implemented for training, so I found my options were very limited using it for training my own data (Transfer Learning) or do much beyond running Tensorflow protobuf models. I started wrapping this library a while back when I realised it was a bit of a deadend for doing much else.
This is still important though, general (and fast) loading and running module for Tensorflow models would be very useful in many circumstances, but the drawbacks and short comings led me to develop VLML.
Lastly, double check if Yolo hasn’t been implemenetation in VL.OpenCV yet.