# How it works

This page explains how AR 51 turns camera video into usable 3D motion — the data path from the cameras to your application. AR 51 is **markerless**: nothing is worn by the performer. The pipeline has three stages — **Capture → Compute → Consume**.

<ThemedImage
  alt="The AR 51 markerless pipeline: Capture with 9 MP / 120 FPS cameras, Compute on the GPU vision server, Consume in Mocap Studio and game engines"
  sources={{light: pipelineLight, dark: pipelineDark}}
  style={{width: "100%", height: "auto", margin: "1rem 0"}}
/>

## 1. Capture

AR 51's MindVision cameras (**9 MP, 120 FPS**, higher frame rates supported) ring the capture volume and stream **synchronized** video to the server. The cameras are hardware-synced so every frame across the array shares a timestamp — this is what lets the next stage fuse views correctly.

→ See [hardware overview](/docs/hardware/hardware-overview) and [room & camera setup](/docs/physical-setup/camera-mounting).

## 2. Compute

The **computer-vision server (CVS)** runs GPU pose estimation, fusing all camera views into 3D data many times per second. Per frame it produces:

- **Skeletons and hands** for every tracked person
- **Tracked objects** you've registered (props, tools)
- **Camera poses** from calibration, so all output shares one coordinate space

It tracks multiple people simultaneously and re-identifies them across frames via a persistent `EntityId`, so a person keeps their identity after leaving and re-entering the volume.

→ Fusion depends on a one-time [camera calibration](/docs/calibration/camera-calibration); identity handling is covered in [entity identification](/docs/mocap-studio/entity-identification).

## 3. Consume

The 3D output is consumed in two ways.

**Mocap Studio** — visualize the capture, record takes, and export (FBX and other formats).

**SDKs and APIs** — stream the data live into your own application over **gRPC**. Available clients:

| SDK / API | Language | Typical use |
|---|---|---|
| [Unity SDK](/docs/unity-sdk/overview) | C# | Unity games/apps, VR, virtual production |
| [Unreal SDK](/docs/unreal-sdk/overview) | C++ / Blueprint | Unreal projects, LiveLink, RenderStream |
| .NET | C# | Headless / desktop consumers |
| C++ | C++ | Native integrations |
| Python (PyCvs) | Python | Research, data pipelines, ML |

Clients don't hard-code addresses: they discover services through the **OMS** registry and connect. See Connecting a client.

## The pieces, in a sentence each

| Term | What it is |
|------|------------|
| **CVS** | Computer-vision server — runs pose estimation and produces the 3D motion from camera video. |
| **OMS** | The registration/discovery service that lets components find each other. |
| **DGS** | Shared scene & spatial anchors for multi-user / VR sessions. |
| **EntityId / PersonId** | Persistent vs. per-session identities for tracked people. |

Full definitions in the [glossary](/docs/introduction/concepts-glossary).

## Where to go next

- [Quickstart](/docs/getting-started/quickstart) — go from a running system to your first capture.
- [SDK & API → Architecture](/docs/sdk-api/overview) — service topology and the data model.
- Connecting a client — discover services and open a stream.