Lyres AI Bottle Recognition - Impossible Bar

Unbnd were augmenting Lyres product range for the 'Impossible Bar' campaign when they found that the AR library they were using couldn't fit enough cylindrical markers in to augment the whole range, nor could the library distinguish between the 750ml and 250ml bottles of the same spirit.

I was brought in to train and implement a custom ML model to categorise bottles by spirit and size. In the process I was careful to cultivate a varied dataset (captured on a variety of mobile devices that might use the model, and in a variety of lighting/backgrounds), and minimise the model size for web deployment.

The biggest concern was how viable differentiating between a 250ml and 750ml bottle at close range was. Thankfully the final model had a one-shot accuracy of 90%, and when I fed the results through a multi-frame confidence filter, we reached 100% real-world accuracy.