Hi, so this Pi has TensorFlow and protobuf built and installed. First I will show it using label_image.py from https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/label_image to classify a panda. Yes, I know. In this case, we use a graph file from InceptionV3 off-shelf.  

Next, I will use the label_image.py code from https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/image_retraining/label_image.py and I'll pass the my graph file from retraining InceptionV3 to this, along with the labels file. See here for the retraining instructions: https://hackaday.io/project/20448-elephant-ai/log/68436-software-elephant-detector-transfer-learning-off-shelf-model-with-tensorflow

My retraining on InceptionV3 was to add two new classes: herd elephants, and lone elephants. So we can see what kind of results we get for classification of images containing herds of elephants, and lone elephants!

It takes around 10-15 seconds for the Raspberry Pi to classify the images. That's with 16MB to GPU, and running a GUI. It should be faster if you don't use a GUI. I'll try and get some comparisons later.

Here's the video! 

I've uploaded the graph file here <add it>

And the labels.txt here <add it>

So you can try! However, this is just from a test retraining run, and I only used 50-100 images per class.