Many cameras in,
one motion out
Synchronised camera views fuse into a single calibrated 3D skeleton. The output is recoverable from any angle - the centre clip here is such reconstruction, with the virtual camera rotating around it.
Calibrated.
Not magic
Multi-view geometry is a sixty-year-old discipline. We use it correctly — known baselines, DLT triangulation, reprojection-error gating, audit-trail outputs.
- Linear DLTN-view triangulation, confidence-gated
- Pinhole-from-geometryMultiple source capture · chessboard planned
15 mm of mean
joint error
Four cameras around a dance sequence on the public AIST++ benchmark — no gymnast, no worker, no domain-specific tuning. The geometry is sport-agnostic by construction, and every number is reproducible end-to-end from the same artefacts our pipeline writes.
Judging-assist for rhythmic gymnastics
Two cameras for training-grade capture in clubs and academies; scale to four or more with an optional LiDAR module for competition-grade precision and apparatus-touch detection. Every metric ships with a confidence interval and the frame range a judge can replay.
Gymnastics is
where we start
The same calibrated stack measures any motion worth measuring. The geometry does not care whether you put a gymnast, a worker, or an inspector in front of it.
- —Occupational ergonomics
- —Protocol compliance
- —Healthcare procedures
- —Other sports
- —Industrial maintenance