Record with multiple cameras
A single camera can produce a recording, but multiple cameras produce better ones: more viewpoints mean more information about where a joint actually sits in 3D, and a lot less vulnerability to one camera's view getting blocked. This page covers how many cameras to use and where to put them; for the physical setup itself (cables, lighting, background), see Optimize your capture space, and for linking your cameras into one coordinate system, see Calibrate your cameras.
How many cameras do you need
There's no universal answer, and published guidance disagrees: theia3D requires a minimum of six cameras and suggests at least eight, while OpenCap studies have found only minimal accuracy benefit moving from two cameras to five, and other OpenPose-based systems have found substantial improvement from adding cameras. FreeMoCap's own validation work used six cameras (mostly because that's how many USB ports were available on the lab computer), but got accurate overground gait results with just three.
As a starting point: use at least three cameras, and add more if your task involves a lot of self-occlusion (one body part blocking another) or covers a wide physical area. The right number is task-dependent, not a fixed rule.
Positioning for a fixed space
If your subject stays roughly in one place (a treadmill, a chair, a small capture volume), surround them with cameras rather than lining them up on one side. A few things worth knowing:
- Purely front-on or purely side-on (sagittal) views tend to invite more occlusion and give the reconstruction less to work with; oblique angles between cameras generally do better.
- That said, this isn't an absolute rule. Equipment that occludes a front-on view (like a rowing machine) can make a purely sagittal setup the better choice for that specific task. Treat positioning guidance as a starting point to adjust from, not a law.
- Whatever the layout, make sure each camera shares an overlapping view with at least one other camera, both for reconstruction and because calibration depends on cameras seeing the ChArUco board together.
Positioning for a moving subject
Overground walking or any recording where the subject covers real distance needs a different strategy. Instead of framing the subject tightly, cover the space they'll move through. One effective pattern uses two cameras in portrait orientation at opposite ends of the walkway, angled to look down its length, plus several cameras in landscape orientation along the side for lateral coverage. Mixing portrait and landscape orientations like this maximizes ground coverage in a way that an all-landscape layout doesn't. The goal, wherever the subject ends up along the path, is at least two cameras seeing them at once.
When the "rules" don't apply
Camera positioning is genuinely task-dependent, and the guidance preceding comes from validated setups (treadmill gait, overground gait, a stationary rowing machine), not from an underlying formula. Expect to iterate. Record a test clip, check whether reconstruction quality holds up across the movement you actually care about, and adjust from there.