Framework · TangibleTable · Product
Custom C# Neural Network for Tangible Marker Recognition
A custom-built neural network engine developed from scratch in C# within Unity to solve tracking instability and jitter in physical tangible table markers.
- Client
- TheHangingHouse (Codebrew)
- Platform
- Desktop
- Role
- Solo
A machine learning solution engineered for a tangible table interface utilizing physical markers with three touch points forming unique triangles. Due to minor movement jitter and physical shifting on the table surface, geometric triangle ratios fluctuated slightly, causing recognition failures in the legacy system. To overcome this, a complete neural network framework—including matrix operations, forward prediction, training routines, and weight serialization/deserialization—was built entirely from scratch in C# without external ML libraries. A custom data-collection application was also developed to gather labeled motion data, successfully training the network to recognize markers under movement variance with over 98% accuracy.
Highlights
- This was the first time I have ever tried and learned about a neural network.. its my first experiment in them.. And at that time I gave my promise and was able to build the network completely from scratch in Unity and CSharp, I didn't like to go to another programming languages or trying new things.. I always developed things in CSharp.. that's why when ever I saw python.. I aways discard it and I really wanted to build it from scratch.. I read a full books, spent hours and hours.. I remember my eye getting big like a ball from the dehydration and from focusing on reading and trying things.. On incident I still remember until now, when everything was working.. the math and the gradients and everything.. but the neural network is not training.. I spent I think like around one week and it's not working.. that was the time I realized changing the learning rate could solve it..
- It was really nice challenge, and I learned alot from it.. later on we have used this activation so many times.. and it was really disserve all the hard work.
