When Gollum snarls, when a football video game's players jog exactly like real athletes, when a crowd of digital soldiers move with unrepeatable human sloppiness — nobody keyframed that by hand. A performer did it, wearing a suit dotted with markers or bristling with sensors, and a motion-capture ("mocap") system turned the performance into animation data: a stream of numbers, one sample every few milliseconds, describing where every body part was.

This page is about the machinery that does the recording. There are several rival technologies — optical cameras, inertial sensors, and pure computer vision — and each solves the same problem with a completely different trade-off between accuracy, occlusion, portability and cost. Understanding those trade-offs is the whole game, because the failure modes (an optical marker vanishing behind an arm; an inertial sensor slowly drifting away from reality) are exactly what a downstream animator has to clean up. This lesson feeds the rest of the mocap strand, where we retarget, filter and blend that raw data.

What comes out: a time-series of poses

Marker positions and joint orientations are two views of the same performance. Positions are what a camera literally sees; orientations are what a skeleton needs. A solve step (fitting a skeleton to the marker cloud) converts one into the other — but that comes later. First, how do we get the numbers at all?

Optical, marker-based: the studio workhorse

The classic high-end setup is passive optical. The performer wears small retroreflective markers — little spheres coated to bounce light straight back toward its source. Each camera has a ring of infrared LEDs around its lens; the markers blaze back as bright dots against a dark background, trivially easy to find in the image. Many cameras (often 8–50+) ring the volume and are synchronized to the microsecond, all shuttering at the same instants.

Each camera only sees a marker as a direction — a 2D dot, which back in 3D space is a ray from the camera's focal point outward. One camera can't tell how far along that ray the marker sits. But if two calibrated cameras both see the same marker, we have two rays, and the marker must lie where they cross. Recovering the 3D point from multiple rays is called triangulation, and with well-calibrated cameras it is astonishingly precise — sub-millimetre. The price is occlusion: a marker hidden from view (an arm swinging across the chest) simply disappears, leaving a gap.

The geometry of triangulation

Below, two calibrated cameras each see a marker as a single direction. Camera A's ray and camera B's ray both pass through the marker, so the marker sits at their intersection — the one 3D point consistent with both views. Read it as a piece of pure geometry: two lines in a plane, generally meeting at exactly one point.

Now imagine camera A is blocked — the performer's own shoulder gets in the way. Ray A is gone, and a single ray from B is not enough: the marker could be anywhere along B's line. This is exactly why studios use so many cameras. A third camera, viewing from a different angle, very likely still sees the marker; its ray crosses B's, and the point is recovered despite A's loss. More cameras means more chances that some pair has a clear line of sight at every instant.

Worked example: crossing two rays

Work in a horizontal slice so we can use 2D coordinates (metres). Camera A sits at (0, 0) and reports the marker along a ray heading up-and-right at 45^\circ — direction (1, 1), so the ray is the set of points (t, t) for t \ge 0. Camera B sits at (4, 0) and reports the marker up-and-left at 135^\circ — direction (-1, 1), giving points (4 - s,\, s).

The marker is where the rays meet, so set the coordinates equal:

Substituting t = s into the first equation gives s = 4 - s, so s = 2 and t = 2. Both rays land at (2, 2) — the marker's recovered position. Two cameras, two rays, one intersection.

If camera A is occluded we keep only (4 - s,\, s): an entire line of possible marker positions, with no way to choose among them. A single ray fixes a direction, never a point — hence the need for a second (and, against occlusion, a third) camera whose ray pins down where along B's line the marker truly is.

Active optical, and going markerless

Active optical systems replace passive reflectors with pulsed LEDs on the body. Because each LED can blink its own on/off pattern, the system knows which marker is which without guessing — markers are auto-labelled, which saves a lot of cleanup where passive markers get swapped or confused. The trade is added weight, batteries and wiring on the performer.

Markerless capture drops the suit entirely. A multi-view markerless rig uses several ordinary cameras plus computer vision to infer the body's pose directly from the video. At the cheap, flexible extreme, a single camera with a machine-learning pose estimator can lift a rough 3D skeleton out of ordinary footage — no studio, no markers, even from a phone. The catch is accuracy: markerless output is far noisier and less precise than marker-based triangulation, and it struggles with self-occlusion and unusual poses. It trades fidelity for reach.

Inertial: cameras optional

An inertial suit takes a completely different route: no cameras at all. Each limb segment carries an IMU (inertial measurement unit) — a tiny sensor package holding an accelerometer (senses linear acceleration and, at rest, the direction of gravity), a gyroscope (senses rate of rotation), and a magnetometer (senses the Earth's field, a compass for heading). Fusing the three gives each segment's orientation in space, and stitching the segments together via the known skeleton gives a full pose.

Because there is nothing to see and nothing to block, inertial capture is immune to occlusion, works outdoors and on location, and is genuinely portable — you can capture an actor sprinting across a field. Its weakness is drift: orientations (and any position estimated by integrating acceleration twice) accumulate small sensor errors over time, so the captured figure slowly wanders away from where it really is. Magnetometers also get confused near steel and motors. Inertial buys freedom of movement at the cost of absolute position.

Comparing the technologies

No single system wins on every axis. Roughly:

The two dominant failure modes — optical occlusion and inertial drift — are complementary, which is why serious productions increasingly run hybrid rigs: inertial sensors carry the pose through occluded moments, while occasional optical fixes pin the drifting figure back to absolute position. Each covers the other's weakness.