Absolute location priors in rotation averaging, global positioning, and BA

[sarlinpe]

Summary

Anchor the reconstruction to an external world frame using absolute localizations of some frames, e.g. from visual localization against a georeferenced map. The reconstruction is then expressed in that frame and the priors make it robust to drift.

Input (--location-priors <file.npz>, format documented in vidmap/mapper/location_priors.py): for the localized frames, image names, a confidence in [0, 1], R_cam_from_world with a 6x6 pose covariance, and 2D-3D correspondences (CSR) to a set of world points. Use --gt-reconstruction with a COLMAP model of the localized frames to force them to be keyframes.

• Rotation averaging: Cauchy-robust absolute rotation priors weighted by their covariance and confidence; the spanning-tree initialization is aligned to the prior frame.
• Global positioning: bearing constraints from the 2D-3D correspondences, stiff in GP1 to initialize positions and softer in GP2 so that coherent blocks of outliers can be overruled. GP1 is aligned to the priors with a 4-DoF (scale, translation) RANSAC, separately for each part of the reconstruction that is only weakly covisible with the rest (e.g. across cuts, counting only the tracks that GP1 explains), with translation only when the anchors are too close to observe the scale; the scale prior is relaxed once at least two frames are anchored.
• Bundle adjustment: robust reprojection constraints to the prior world points; the first pose is no longer fixed when priors are active.
• Native costs: the residuals are appended to the upstream COLMAP problems: rotation priors (on the Eigen-quaternion frame rotations, which also frees the averager's fixed frame) in a new vidmap_native.rotation_averaging module, bearing constraints in vidmap_native.global_positioning, and constant-point reprojections in vidmap_native.bundle_adjustment.
• Keyframes: keyframes.selection.force_gt_keyframes (enabled by --gt-reconstruction).
• CLI: --location-priors in vidmap.run and vidmap.map; vidmap.map also honors --split-cuts.
• Viewer: reference (GT) poses aligned by Sim(3), with optional per-frame confidence colors and position covariance ellipses read from pose_covariances.npz / pose_confidence.json next to the reference reconstruction, plus hover tooltips.
• Tests for the npz loader, the rotation-averaging alignment to the priors, and the GP/BA costs.