Digital Twins
Design & Technical Process
How can the intersection of augmented reality, computer vision and intelligence affect our fundamental understanding of space?
A Third Eye
Just like the Meta Ray-bans or EvenReality's glasses, they fundamentally change how we observe and inhabit our environments because of the new ways these devices have allowed us to interact with the world. Our lives may have a dramatic increase in quality when we fully optimize for computers registering space like we do.
Intentional Assembly
Assembling disparate datasets together is an intentional craft, and requires the collaboration of front-end designers and engineers to understand who and what the trained model ultimately serves. This requires a multi-disciplinary approach throughout the whole process of collecting data to deploying the model.
Key Takeaways
A well-trained model begins with high quality datasets and real-world testing to achieve reliable spatial intelligence.
Model Optimization
A model deployed on AR devices should be lightweight enough to conduct faster, real-time inferences, which require less computational power. This ultimately reduces latency and eliminates the need for cloud-based inferences.
This consideration, though, needs to be balanced with accuracy in object-identification recall. The graph below shows a (mostly) successful, lightweight model that has achieved precision-recall above 85%. However, the dataset for "rings" has underperformed due to the lack of training images compared to the rest of the other classes.
