cross-camera-metric-calibration

Use when relating two cameras' views metrically (projecting one camera's 3D into another, borrowed-depth supervision, multi-view rigs) — lock consumer-camera auto modes, checkerboard-calibrate intrins

SKILL.md

GitHub

When to use. Any two-camera (or camera+sensor) metric relationship. Why: every failure mode here was silent — spec-sheet FOVs that match no runtime mode, autofocus drifting the focal length, PnP that is right only up-to-scale, and detector labels that alias across views.

Procedure.

  1. Kill the auto modes first (auto-framing, autofocus, FOV) in the vendor tool; recapture working frames immediately before calibrating. Measure the printed checkerboard square — printers rescale.
  2. Calibrate intrinsics per camera (findChessboardCornersSB for low-contrast boards); a spec FOV is a candidate, never a calibration. Single-view PnP without calibrated intrinsics is scale-ambiguous.
  3. Solve the extrinsic: 3D points from the metric camera (stereo/depth) ↔ 2D features in the other; seeded solvePnPRansac + solvePnPRefineLM. Record which physical act invalidates what (re-aiming kills extrinsics; intrinsics survive if FOV/focus untouched).
  4. Validate end-to-end, class-agnostically: project detections across cameras and match by geometry (center distance / box containment) — labels alias across views (the same figurine read "person" and "teddy bear"). Save the overlay as evidence.
  5. Done when: feature reprojection 1 px, cross-projected items land on their counterparts (10 px), and the invalidation rules are written into the resume surface.

Knowledge: FIELD-KNOWLEDGE-GILBERT.md §G2.