Link Search Menu Expand Document

ossia

AI Recognition

How to run ONNX models in score

AI Recognition

1. Drop an AI model process in your timeline

Let’s try with BlazePose Detector. It can read from a camera or image input texture, run the model, and output results in real-time.

BlazePose Process

2. Download and load model

3. Add an Input camera device (or video)

Choose a camera

Camera input

Alternatively, drag & drop a video file into score, and connect it to the AI model process.

4. Add a window device for output

Adjust the output window size to match your model’s output.

Output window

5. Play

Add a trigger if you want the process to loop continuously.

Trigger example

6. Extract keypoints

Each AI model outputs keypoints in a specific format.
For BlazePose, refer to the official documentation:

BlazePose Keypoints

To extract keypoints like wrist or nose, use the Object Filter process to filter these fields.

  • Left wrist = .keypoints[15].position[]
  • Right wrist = .keypoints[16].position[]
  • Nose = .keypoints[0].position[]

For other models like RTMPose, here is the 26-keypoint layout:

RTMPose keypoints

7. Use keypoints

Once keypoints are extracted, you can connect them to any parameter in score for interactive control.

Use keypoints example

Download score

Example usage of keypoints

You can also send keypoints over OSC to external tools like Wekinator for gesture recognition or AI-based interaction.

  1. Use an Object Filter to extract:
    • left wrist, right wrist, nose (x, y, z) with the following formula : 
      [ .keypoints[16].position[], .keypoints[15].position[], .keypoints[0].position[] ]

  2. Combine them into a 9-value OSC message using the Array Combinor process

  3. Send the OSC message to wekinator:/wek/inputs

OSC to Wekinator

This can be used to trigger visuals, audio, shaders, and other cool things!