Here are some of the ideas I've envisaged for improvements to, and future directions for, the ElephantAI system!

LOCAL PEOPLE TAKE PART IN SUPERVISED LEARNING OF THE ELEPHANT DETECTOR MODEL

In a more advanced PPP (via 3G/4G) connectivity scenario I have envisaged, the input of local people could be leveraged to improve the accuracy of our elephant detector by enabled a supervised training system. Here's what it would look like:

1. Elephant detector tweets/DMs image of suspected elephant

2. Local people tweet back with hashtags #yeselephant #noelephant #type_is_lone #type_is_calf
#type_is_herd #type_is_male #type_is_female according to what they identify in the image
3. Database of images containing elephants confirmed by users is created by elephant detection device
4. Images being stored along with their class labels that were provided via the hashtags (e.g. #type_is_calf , #type_is_herd etc.)
5. On a weekly or monthly basis, the elephant detection devices will upload their databases of images to the cloud (e.g. to an Amazon EC2 virtual machine)

6. Transfer learning will be performed on the virtual machine using our existing dataset with these additional images. Thus improving the accuracy of our elephant detector model!

7. The new model will be sent to the elephant detection device from the cloud-based virtual machine, and be used for future detection.

This process will be repeated on a weekly or monthly basis, thus providing on-going improvements in accuracy of the elephant detector!

Note that, even if we don't have good 3G/4G connectivity from the elephant detection devices to perform this function, we could connect them to a upload station. The upload station would be located most likely in the village, were either wired or 3G network coverage is available. The elephant detection devices can be linked to the upload station via RF comms, either directly or via repeater stations. A similar concept to that I'd envisaged in my first https://hackaday.io/project/10391-automated-elephant-detection-system !