Climate Change AI Hackathon
- We chose to build a plant disease identification model and deployed it on an android app.
- We thought about how the app could be useful to farmers and realised that it would make sense to add the ability to be able to record the predictions made by the app and serve it to a monitoring application. This would enable to them to keep track of how the disease progresses over days. The dashboard shows the location of where the disease images were taken (GPS coordinates and the corresponding predictions along with the timestamps).
- We think the spatial localization is much more important than just using an app and getting predictions from a model. Farm land usually spans a large area with multiple crops and hence keeping track of where the disease plants were observed can be quite useful.
- We faced many technical challenges but were able to hack quick and simple solutions to demonstrate that the concept could work in the real world.
- Deployed on mobile app running classification at 25 frames per second on Android Snapdragon GPU
- GPS, timestamp and image tagging for triggered classes on a live web-app
- creates 3D volumetric sparse map and tracks camera position in real-time - for more accurate diseased plant localization than GPS. This also works in environments where there's no GPS signal