BodyPixSentis is an implementation of the BodyPix model for person segmentation and pose estimation. It runs on the Unity Inference Engine, Unity’s official neural network runtime optimized for real-time applications.
- Unity 6
- Compute shader support
The original BodyPix model, provided in TensorFlow.js format, has been converted to ONNX using tfjs-to-tf and tf2onnx. For details, see the Colab notebook.
This package supports ResNet-based models, which offer higher accuracy but are larger and slower. Due to GitHub and npm.js file size limits, these ONNX files are not included in the repository. You can download them from ResNetZip instead.
To use ResNet models, create a new BodyPix ResourceSet and set the model, architecture, and stride fields accordingly.
You can install the BodyPixSentis package (jp.keijiro.bodypix) via the
"Keijiro" scoped registry using the Unity Package Manager. To add the registry
to your project, follow these instructions.