Motion Input Grid (MIG) Python package

English | Français

Install from PyPI

python -m pip install motion-input-grid
python -c "from mig import Tracker; print('MIG OK')"

The distribution name is motion-input-grid; the import remains mig. Prefer a Python 3.10+ virtual environment. Wheels cover Windows x64 and Linux x64/ARM64 with glibc 2.35+. On other platforms, pip may build the sdist, requiring a C++20 compiler and CMake 3.25+. To select a version: python -m pip install motion-input-grid==1.1.0.

Local alternative: python -m pip install /path/to/motion_input_grid-1.1.0-<tags>.whl. To build from a full checkout: python -m pip install ./bindings/python.

Submit observations

Save a Motion Input Grid Studio profile as profile.json. Pass None to select the bundled engine or an explicit path to an external SDK. This example submits one frame; your application should feed the tracker each new observation. One frame alone does not calibrate or complete a movement.

Python accesses the C++ engine through standard-library ctypes; no Python inference dependency is required. Wheels include the positions C ABI and licenses: mig-c.dll on Windows, libmig-c.so.1 on Linux. No camera estimator or models are included. The sdist contains the canonical engine and JSON headers; rebuilding requires CMake 3.25+ and a C++20 compiler. Linux builds use GCC and its runtime copyright notices.

from pathlib import Path
from mig import Packet, Tracker

profile = Path("profile.json").read_text(encoding="utf-8")
with Tracker(None, profile) as tracker:
    frame = Packet(timestamp_ms=20, sequence=1, aspect=16 / 9)
    frame.set_body(11, 0.65, 0.45)
    frame.set_body(12, 0.35, 0.45)
    frame.set_body(15, 0.4, 0.6)
    for action, input_id in tracker.update(frame):
        print(action, input_id)

Lifecycle and camera

Use increasing timestamps in milliseconds and increasing sequence numbers. Calibration requires roughly one second with both shoulders visible. Positions use unmirrored anatomical MediaPipe indices; missing points have zero confidence. Keep each tracker on one owning thread and close it explicitly.

start_camera(runtime_directory) and poll_camera() optionally use a native camera-enabled SDK plus its MediaPipe runtime and models. Poll already updates recognition; consume events() afterward without submitting the frame twice. Action callbacks are logical events; keyboard injection is the host's choice.

See the repository's examples and ABI contract for body and hand XYZ packets, capture ownership, Pygame and Tkinter integration.

Complete source walkthrough · Bootstrap.