"Was ist das?" — Object Recognition App for Children
Problem
Children who can't read yet have no independent way to find out what an everyday object is called — they have to ask an adult. I wanted a tool a young child could use completely on their own.
What I Built
An offline iOS app: point the camera at an object, and the app recognizes it and speaks a short, child-friendly answer out loud. The model recognizes 48 object classes. Training data was prepared with Python scripts pulling from the Open Images Dataset v7 and COCO, and the classifier was trained with Core ML and Create ML. Spoken answers are pre-generated with ElevenLabs so the app never needs a network connection at runtime.
Technical Decisions
The core constraint was that this is a product for young children: it had to work fully offline, with no data ever leaving the device — no analytics, no cloud inference, nothing to explain to a parent. That ruled out server-side vision models and shaped the whole pipeline around an on-device Core ML model instead. The model reached about 85% validation accuracy against about 70% training accuracy.
Result
Delivered as my Bachelor's thesis (HTW Berlin, 01/2026–03/2026). A working, fully offline object-recognition app with a trained 48-class model and natural-sounding spoken output, built solo from data collection through to the final iOS app.