I tried adding another 500 carefully crafted images to the database and ran the training session once more, but got no improvement in mAP and no noticeable improvement in a real life deployment on the Jetson TX2. According to the forums etc. its normally all about getting a larger dataset, but in my case this has made no difference at all!

Maybe it's time to start hacking into the standard bvlc_googlenet.caffemodel pre-trained model?

Does anybody have any suggestions?