Preliminary study
The scene, what a camera sees, recognising the T, the bundle adjustment and synchronisation at 30–120 fps.
Transcript
1. The scene
8 cameras around a 3.5 × 3.5 m volume. A person carries the T for 15 s; the highlighted camera is the one we follow.
2. What a camera sees
Only nameless 2D points: sometimes a marker is hidden by the body, sometimes a reflection shows up.
3. Recognising the T
The affine fit tries every assignment and discards ambiguous frames. Compare with the symmetric T: every frame is ambiguous.
4. First estimate
The static terna gives a first position of the cameras. Here the start is deliberately worse: 8°, 50 cm, 15% focal.
5. Bundle adjustment
Cameras, focal lengths, distortion, delays and T positions are corrected together. The reprojection goes down; the true 3D error oscillates, then settles at 1.5 mm.
6. Synchronisation
Two cameras, a foot at 6 m/s. Pick the fps and the mode: the offset between shots becomes an error on the foot.
7. Result
With the asymmetric T, multi-view relabelling and the estimated delay: about 1 mm over the covered volume.