This repo is only for demonstrating Nandi platform. Development codes are kept confidential.
This project is aimed to develop a platform for
- Enabling Digital Identitiy of livestock to enable tracing and logging of everyday data.
- Performing Valuation based on body condition scoring from parameters extracted using images of the animals.
Currently, eartags are used for logging daily data and registering livestock in government database. Once the tag gets ruptured or thrown away, all the collected data is lost. Our platform utilizes images of animals and creates a digital identity for them using their own intrinsic features. With this project, we are developing largest available cattle database containing about 400 million dairy animals.
Currently, further field validation is going on for the biometrics part and Data collection phase is ongoing for cattle valutaion part.
- Completed Pilot phase for Government of India with data collection of more than 1000 animals from 4 states and 12 regions across India, including both cows and buffaloes.
- Demonstrated Few shot learning capabilities with 92% on-field top-1 identification accuracy on cross breed cattle, indigenous cows and buffaloes.
Entire stack consists of an android app, an admin webapp and a backend Restful API which connects everything together. Everything is bundled into docker containers and deployed on Azure Cloud.
An android app is built for users which has a .tflite quality check model deployed on it. This model checks for several parameters for the quality of image and only lets through the images which pass the QC. Using the app, users can register new animals as well as perform identifications on previously onboarded animals.
Admin webapp is built using wordpress. Admins can monitor incoming data, verification logs and analytics graphs based on metadata collected.
This api is built using FastAPI and holds the entire stack together. It stores the incoming data from the app into the Azure blob storage and metadata to mongoDB. It runs all the models and retrieves the verification results and sends it to the app as well.
Verification/Identification of the animal works as follows
- User clicks an image of the animal.
- QC model checks for the quality parameters and either accept or reject the image. (Notice green tick in the video below)
- If image passes QC, user submits the image else start from step 1 again.
- QC pass image is sent to backend along with metadata.
- Backend generates embeddings for this image and filters the identification search space based on metadata collected.
- Search results (top 3 matches) are sent back to the app.
- Feedback from the user is recorded to further tune the model.