vivid-onnx

Machine learning inference for creative applications via ONNX Runtime.

Installation

vivid modules install https://github.com/seethroughlab/vivid-onnx

Operators

Operator	Description
ONNXModel	Generic ONNX model inference
PoseDetector	Body pose detection using MoveNet

Included Models

The addon includes pre-trained models in assets/models/:

Model	File	Description
MoveNet SinglePose Lightning	movenet/singlepose-lightning.onnx	Fast single-person pose (9MB)
MoveNet MultiPose Lightning	movenet/multipose-lightning.onnx	Multi-person pose (19MB)

Models from PINTO_model_zoo.

Examples

Example	Description
pose-tracking	Real-time body pose with skeleton overlay

Quick Start: Pose Detection

#include <vivid/vivid.h>
#include <vivid/video/video.h>
#include <vivid/ml/ml.h>
#include <vivid/effects/effects.h>
#include <cstdlib>

using namespace vivid::video;
using namespace vivid::ml;
using namespace vivid::effects;

void setup(Context& ctx) {
    auto& chain = ctx.chain();

    // Webcam input
    chain.add<Webcam>("webcam")
        .setResolution(1280, 720);

    // Pose detection - model path for installed library
    std::string home = std::getenv("HOME") ? std::getenv("HOME") : "";
    chain.add<PoseDetector>("pose")
        .input("webcam")
        .model(home + "/.vivid/modules/vivid-onnx/src/assets/models/movenet/singlepose-lightning.onnx")
        .confidenceThreshold(0.3f);

    // Canvas for skeleton overlay
    chain.add<Canvas>("skeleton")
        .size(1280, 720);

    chain.add<Composite>("output")
        .input(0, "webcam")
        .input(1, "skeleton")
        .mode(BlendMode::Over);

    chain.output("output");
}

void update(Context& ctx) {
    auto& pose = ctx.chain().get<PoseDetector>("pose");
    auto& canvas = ctx.chain().get<Canvas>("skeleton");

    canvas.clear(0, 0, 0, 0);

    if (pose.detected()) {
        // Draw keypoints using Canvas path API
        for (int i = 0; i < 17; i++) {
            auto kp = static_cast<Keypoint>(i);
            if (pose.confidence(kp) > 0.3f) {
                auto p = pose.keypoint(kp);
                canvas.fillStyle(1, 0, 0, 1);
                canvas.beginPath();
                canvas.arc(p.x * 1280, p.y * 720, 8.0f, 0, 6.28f);
                canvas.fill();
            }
        }
    }
}

VIVID_CHAIN(setup, update)

Keypoint Indices (MoveNet)

Index	Keypoint
0	Nose
1-2	Left/Right Eye
3-4	Left/Right Ear
5-6	Left/Right Shoulder
7-8	Left/Right Elbow
9-10	Left/Right Wrist
11-12	Left/Right Hip
13-14	Left/Right Knee
15-16	Left/Right Ankle

API Reference

See LLM-REFERENCE.md for complete operator documentation.

Dependencies

• vivid-core
• vivid-video (for Webcam input)
• ONNX Runtime

License

MIT (addon code) Apache 2.0 (MoveNet models)