Hi, I am attempting this as a proof of concept for identification of individual animals using retrained CNs (i.e. TensorFlow with Inception V3). The retraining being done on a machine with a GPU, the classification using the graph file being done on a Raspberry Pi. It could have useful applications in access control for cat doors and feeding stations. In addition to cat-tracking to declared locations (e..g did cat a enter into a certain garden or house/ is cat b in a certain room). Although it is perhaps slightly over-complicated, when RFID can do the same!
I'll just use Raspberry Pi with PIR and camera module to obtain images of suspected cats! We'll just do daytime (IR-filtered camera) detection to begin. I'll start by trying at the cat feeding station, and putting cardboard around, so humans going past don't trigger the PIR. This will be the cat detection device.
1. Put the cat detection raspberry pi in place, and collect images of my cats (they are called LaLa and Po). I'll just power it from AC-DC converter since there is a AC socket nearby.
2. Additionally collect images of the cats whenever possible via iPhones! I'll retrain Inception V3 final layer three times: once with images only collected via the cat detection raspberry pi, once with images obtained via iPhones, and once with both datasets. Since I would like to see if using the same camera and same kind of poses we will get (i.e. the cat detection raspberry pi) will increase accuracy of the retrained model.
3. Go ahead with retraining Inception V3.
4. Deploy the graph file to the cat detection raspberry pi, and get it now to pass images to the image labelling function. After which if it gets "LaLa" label it will go ahead and save that image to LaLa directory. If it got "Po" label it will go ahead and save that image to Po directory.