Multi-camera 3D fusion
Sub-pixel tracking across an uncalibrated three-camera rig
A production capture setup needed frame-accurate 3D positions for small, fast-moving marked objects — across three cameras at wide baselines, one of them at a punishing oblique angle where detections fragmented into anonymous clusters.
- Built a marker-level fusion pipeline that triangulates each fiducial independently rather than solving per-view pose, eliminating the inter-view "breathing" that board-level solving introduced.
- Cross-view epipolar transfer raised usable marker identification in the worst camera from under 2% to 94–97%, at 100% precision with zero identity conflicts.
- Diagnosed a physical rig defect through leave-one-view-out validation — a fault invisible to closure-error metrics, which reported 0.09°.
- Established per-take epipolar synchronisation after proving that audio and timecode alignment were insufficient at the required precision.
reprojection 0.7–1.3 px
identified markers 94–97%
false positives 0
views 3