The video above shows a common European wasp being 'inferred', which means it's being recognised and it's position and size represented by a blue box. Here, the coordinates of the boxes are exported via I2C to an Arduino where they are processed to produced beeps of various frequencies depending on the confidence, or accuracy, statistic. The higher the beep frequency, the better the accuracy.

The model was trained on my Nvidia Jetson TX2 using caffe with a pre-trained model based on googlenet, the 'bvlc_googlenet.caffemodel'. The 'wrapper' software used is DIGITS 6.1 from Nvidia, which enables an easy to use GUI on a standard internet browser, which is really useful as this system can also be used on a cloud based GPU in the same way.

The TX2 is not the ideal training machine as I could only run 100 epochs with about 2600 images of 640 x 640 pixels and the DIGITS software, for some reason, cant display updated accuracy statistics, which makes it difficult to tell if the training has reached any degree of 'convergence' or not.

The next step is to do more training on the Amazon AWS cloud service which allows us to create an ad-hoc computer in the cloud with very powerful GPU cards for a small per hour fee rather than having to fork out thousands of pounds / dollars for our own system. The AWS system can run Nvidia Docker containers that are already pre-configured with caffe and DIGITS etc so, thankfully, there's no protracted software installation with the usually gauntlet of missing dependencies. Setting up the AWS machine is not, however, trivial and there's still a lot of head scratching involved in getting it to work! More on this in the next update.