Open notes on calibration, triangulation,
and validation

How we calibrate, how we triangulate, how we know we are right. The validation report behind every number on this site — sport-agnostic by construction, applied first to rhythmic gymnastics.

Research — open notes
Latest benchmark
15mm
Mean joint error across the 17 COCO joints on AIST++. Stock pose, linear DLT triangulation, no sport-specific tuning yet — every number reproducible from artefacts in the repo.
AIST++ · 4-cam · linear DLT
Fig.Headline benchmarkoutput/aist_validation.json
Method

Multi-view geometry,
done correctly

Per-view pose, multi-view triangulation, calibration as the dominant accuracy source. Stock components, audited at every step.

Per-view 2D pose
YOLOv8x-pose, COCO-17 skeleton. Stock weights, no sport-specific fine-tuning yet.
Calibration
Pinhole-from-geometry on the gymnast capture path. Chessboard / charuco planned for v0.4.
Triangulation
Linear DLT, N-view, gated by per-joint reprojection error.
Validation
AIST++ ground truth. MPJPE in millimetres, reproducible end-to-end from the repo.
Benchmark

Where the error lives

All joints
< 30mm
Every COCO joint falls below the 30 mm ceiling. Wrists land worst at roughly 30 mm; face and lower body sit 8–13 mm.
Triangulation rate
100%
Every frame recovered all 17 joints. No skip from insufficient view coverage on this clip.
Reprojection error
2–5px
Median reprojection error across all views and joints — the signal the pipeline uses to gate each triangulation.

gLH_sBM_cAll_d17 · 4 cam at 90° · YOLOv8 pose · linear DLT

20 mm — review threshold
noseeye·leye·rear·lear·rshldr·lshldr·relb·lelb·rwrst·lwrst·rhip·lhip·rknee·lknee·rankl·lankl·r
Fig. Per-joint MPJPE across 17 COCO joints, in millimetres.output/aist_validation.json
Known limits

What we are
not yet good at

Four open weaknesses we currently track. Each has a mitigation path in flight and a target version for the fix.

Known limit
Apparatus is the harder half
Ribbon, rope, and clubs deform frame to frame and stock pose models don't see them. Apparatus tracking ships separately, sport-specific, and isn't validated to the same number as the body skeleton yet.
Known limit
Joints disappear under apparatus
Ribbon, hoop, and ball cross the body during throws and catches, briefly hiding wrists and elbows from one or more views. Per-joint confidence flags these frames as needs-review; apparatus-aware fusion improves recovery once it ships.
Known limit
Pose quality varies element to element
Stock COCO-17 weights were trained on everyday motion. Inversions, pivots, and back arches push the 2D model past its training distribution — recovery quality drifts until sport-specific fine-tuning lands.
Known limit
Calibration is the dominant accuracy source
Calibration sets the ceiling. 2× error gap measured between assumed-pinhole geometry and AIST-supplied calibration on identical pipeline — same code, same frames, two times the joint error. Chessboard and charuco calibration in v0.4 close the gap.
Contact
dashavolosova@gmail.com
We reply within five business days