As mentioned in previous logs, there was something seriously wrong with the model trained using mobilenet SSD …… And I don't know what went wrong …. The training and subsequent model optimisation seemed to run smoothly. I ran out of time and patience so have gone back to the Jetson TX2 and made a 'like for like' appraisal using a 'fresh' video from YouTube for testing the detection capabilities. It's important to test the network on footage that was NOT used to train the network.

Using an external USB camera, I could see no false positives on the TX2 and it did not pick up on similar insects such as flies. It's also worth noting that the network was trained using quite a few images with honey bees in the back ground, which will help focus the results away from similar looking insects. 

The bvlc_googlenet model was trained on Amazon Web Services (AWS) for about 5 hours, 63300 iterations with a approx. 2,500 images set, 640 x 640 pixels.