The following list contains recommendations for using Live Sense SDK in your applications.
- Run recognition on a dedicated background thread.
- Recognition can take hundreds of milliseconds depending upon the device and configuration. Avoid calls on the UI thread.
- For running multiple models in parallel, each model must have its own thread.
- If you are facing lag in UI, call the
LSDPixelBufferRecognition?.processPixelBuffer(pixelBuffer: pixelBuffer, orientation: orientation) method from a background thread.
- Use one of the available object trackers to follow objects during intermediate frames.
- Avoid large input images.
- HD (1280 x 720 pixels) is the recommended resolution for recognition of a full frame.
- Start the minimum confidence for recognitions at 0.6 and adjust it based on your project's requirements.
- It is recommended that you not go below 0.6 for your confidence threshold.
- For more information on confidence values, see Models.
- Only include ARM ABIs for release builds to reduce app size, unless x86 support is required for your application.