After abandoning attempts to use C++ as shown in previous log, I reverted to python scripts, which are a lot easier to 'compile'. Hackaday's own Lewin Day pointed towards Les Wilkinson's work with Python and Neural Network Stick 2 and Intel Openvino and after forking Les's repository ( https://github.com/paddygoat/RPi3_NCS2 ) I was quickly able to modify his bottle following robot code to make my gimbal track not only bottles, but also faces and wasps. Having already proven my wasp model to work on openvino, I  now added timers to the python script to track down bottlenecks and found that my model had a fairly large one around the 'inference blob' in the script, which was, to me at least, really interesting!

Using the same script, I was able to demonstrate the difference in performance between two networks, namely 'MobileNet-SSD' and 'Squeezenet 1.0 modified SSD', finding that the older Squeezenet model ran a lot more efficiently on the Raspberry Pi configuration giving higher blob rates and very much better inference accuracy / precision. If the object of interest is relative simple, like a face, it might be worth looking more closely at an older and simpler network such as Squeezenet for getting better results than more complicated networks such as MobileNet SSD.