Frequently Asked Questions
Short answers to the questions people ask most. For more detail, see the rest of the documentation.
Is any of my data uploaded anywhere?
No. mHumanTrackAI performs all capture, inference, gesture recognition and input injection on your own machine. There are no cloud calls, no telemetry and no uploads. The Python engine and the C# app communicate only over the loopback interface (127.0.0.1). The only optional network activity is a one-time model download on first run (from a Hugging Face mirror), which you can also satisfy by placing the model files manually.
Do I need a GPU?
No — the installer detects your hardware and falls back to RTMW on CPU or MediaPipe on any x64 machine. A GPU makes a large difference for the RTMW whole-body model: an RTX 2060 reaches roughly 20–28 fps on CUDA (and ~90 fps with the TensorRT FP16 fast path), versus around 2 fps on CPU.
Does it work with two cameras?
Yes. Adding a second camera (the headline case is front + back) enables multi-view fusion: a per-joint best-view selector chooses whichever camera actually sees each joint, which is how it handles occlusion such as a hand behind your back.
Will it work in competitive games?
Gaming Mode emits a ViGEm virtual Xbox 360 controller, which is accepted far more widely than synthetic mouse/keyboard input. However, kernel-level anti-cheat can still reject synthetic input devices. This is an external limitation that no input tool can fully work around.
What cameras are recommended?
Any USB or built-in webcam works at 640×480 @ 30 fps or better. For the best tracking, a 1280×720 @ 60 fps camera with MJPG support (such as a Logitech C920) is recommended. A second camera is optional and unlocks full-body occlusion coverage.
Do I need to install Python or .NET separately?
No. The release build bundles its own CPython 3.11 and the .NET 8 runtime, so end users do not need to install anything else. Gaming Mode additionally requires the ViGEmBus kernel driver, which the installer can handle.
How does it choose between GPU and CPU?
The single installer detects your hardware automatically. If it finds a CUDA-capable NVIDIA GPU, it uses the CUDA 12 stack (onnxruntime-gpu and NVIDIA runtime wheels) plus an in-app download manager for optional backends such as TensorRT. Otherwise it falls back to RTMW on CPU or MediaPipe, no extra setup needed.
Can I record and replay sessions?
Yes. You can record the live landmark stream to a session file, replay it through the full engine pipeline without a camera (useful for repeatability testing), and export recordings to CSV or BVH for animation tools.