Models

This is a pre-release of two new pose-estimation models. They should generally perform slightly better than the original release. Specifically, note the the following.

• Trained on a larger, more diverse dataset which might transfer to better tracking quality across a wide user base IRL.
• Less twitching when user blinks, due to being trained on semi-synthetic examples with closed eyes.
• Supports the adaptive smoothing filter (built into the neuralnet-tracker) with the help of uncertainty estimation.
• Still unresolved is the pitch-yaw bias / systematic errors.
• Accuracy in terms of angle errors is on par with SOTA literature results from last year.
• The small model should run as fast as the original due to being of the same general architecture.
• The big model is ca 6x slower (on my machine) but less noisy.