Skip to main content

Neural Network Model Preloading

Lightship's awareness features, such as Depth and Semantic Segmentation, rely on neural network models to know how to draw each pixel in an AR environment. Because these models are not included with ARDK, the application must download and install them before the features can provide output, resulting in long load times that can negatively impact the user experience. To prevent this, Lightship provides the Model Preloading feature, allowing applications to download the model beforehand and reduce startup latency when using awareness features. (There will be some latency because of the time that model initialization takes, just not as much as downloading the full model.)

Applications can either register an existing model file or request that ARDK download one. Either way, once preloaded, the model file will remain stored in the application's cache until cleared by uninstallation and will load at runtime.

Known Limitations

  • Model Preloading saves model files to the application cache. Model decryption still occurs when the subsystem starts, so some latency is still expected when starting awareness features.
  • Lightship resets to the default model file definitions at startup and custom model file registrations from previous sessions are not retained. To bypass this limitation, developers can use RegisterModel to re-register model file(s) at the start of each AR session.

More Information

See How to Use Neural Network Model Preloading for details on using this feature in your Lightship project.