Skip to content

Latest commit

 

History

History
32 lines (24 loc) · 1.37 KB

PoseDetection.org

File metadata and controls

32 lines (24 loc) · 1.37 KB
Raw

Pose detection in react-native-camera

This is a fork of the react-native-camera library, modified to allow native pose detection on the preview stream (since passing image frames back to JS over the bridge is too slow).

Most of the changes follow the patterns already implemented for face detection.

Concretely:

  • onModelProcessed is a JS callback that will receive the output of processing each frame.
  • You can pass in a modelParams prop to control the pose detection. This is backed by CameraViewManager.setModelParams, which invokes setModelParams on RNCameraView
  • RNCameraView.setModelParams sets up a TFliteModel (defined in react-native-pose-decoding) to instantiate a TFlite Interpreter. The .tflite model file is expected to be in the assets/ folder.
  • onFramePreview will then start an asynchronous ModelProcessorAsyncTask for each frame, and pass to it a correctly rotated bitmap to process.
  • ModelProcessorAsyncTask delegates most of its work to react-native-pose-decoding’s Decoder . The result of pose detection is a list of Pose objects which are serialized via ModelProcessedEvent back to JS.