Case Study
BeeGuard
In a self-initiated project, SoftTeco built BeeGuard, an AI-powered iOS app that uses computer vision to automatically count and track bee populations.
AI-generated imageCountry
Not specified (internal SoftTeco project)
Industry
Agriculture/Sustainability (Beekeeping)
Platform
Mobile (iOS)
Technologies
Swift5, PyTorch, YOLOv3, Google Cloud Platform, DigitalOcean, Python, Flask
Challenge
BeeGuard wasn’t a client engagement but an internal initiative through which SoftTeco set out to build up its own machine learning expertise. The team chose the task of counting bees in a hive via object recognition and forecasting the population’s development, partly because SoftTeco is a member of the UN Global Compact with a focus on sustainability, and partly because the team included experienced beekeepers. The main challenge lay in building expertise in object recognition, detection and classification, and in learning how to properly collect, process and label the training data needed for the model.
Solution
SoftTeco built the entire solution in-house — from the product architecture (app, server, cloud) through data collection and labeling to training the neural network, which went through more than 2,000 training epochs based on YOLOv3 on the Google Cloud Platform. The trained model was integrated into PyTorch and served via a Flask web server with a purpose-built API for the mobile app. The iOS app itself was built in Swift5 and automatically detects the hive frame and the optimal focus point when a photo is taken, improving image recognition accuracy, while also visualizing inspection results in charts.
Results
SoftTeco successfully shipped a working AI-powered iOS app for beekeepers. According to the case study, the automated photo capture cuts a single hive inspection down to around 10 minutes, compared with up to an hour using the traditional manual method. The team continues to refine the app and plans to bring it to international markets.