A stationary person is a forgiving subject for a multi-camera reconstruction. Move an arm quickly and the system has to reveal how well its sensors agree. A small timing mismatch can create two versions of the same hand, each geometrically plausible in its own frame.
VCL's VolumetricCapture toolset is interesting because it treats capture as a distributed system. Processing units collect data from individual RGB-D sensors, a control application coordinates the system, and a broker carries communication. The project separates spatial alignment from temporal alignment instead of assuming that connecting several cameras makes their observations simultaneous.
Spatial alignment answers where a measurement belongs. A point in one camera's local coordinates is transformed into a shared coordinate system. If that transform is slightly wrong, overlapping observations can form doubled surfaces or visible edges. The disagreement remains even when the subject is still.
Temporal alignment answers when the measurement belongs. Suppose a hand moves one meter per second and two captures differ by 20 milliseconds. The hand travels two centimeters during that interval. No fixed rotation or translation between the cameras can remove that discrepancy for every pose, because the cause is motion between observations.
Clock synchronization and exposure synchronization are related but distinct. Agreeing on the time allows timestamps to be compared. It doesn't necessarily make sensors expose at exactly the same instant. A processing pipeline also needs to know what a timestamp describes: capture, arrival at the host, or some later processing event.
Buffering can align streams by selecting frames with nearby timestamps, but it costs latency. Waiting for the slowest camera may improve the consistency of a recorded volume while making live conversation feel less immediate. Dropping a late frame preserves freshness but can reduce coverage. The application has to choose which failure is easier to tolerate.
Color introduces another seam even when geometry and timing agree. Different exposure or white-balance decisions can make one side of a person visibly change where observations are blended. More cameras provide more coverage, but also more opportunities for incompatible appearance.
I would diagnose these problems separately. A static target helps reveal geometric misalignment. A moving object helps reveal timing errors. A uniform surface under stable light helps expose photometric differences. Trying to fix everything by smoothing the final cloud can hide symptoms while discarding useful detail.
The distributed architecture also makes recording discipline important. Preserving sensor identity, calibration and timing information alongside frames gives later processing a chance to improve the result. A folder of images without those relationships is much less valuable than it appears.
The system is building one account of an event from several witnesses. Each witness has a viewpoint, a clock and a response to light. A convincing volume depends on understanding those differences well enough that they stop appearing as differences in the person being captured.