Auditing ambiguous multiview geometry.

What would make a repair useful?

A controlled repair must improve independently checked correctness while preserving useful views and reconstruction completeness. Removing difficult inputs alone is insufficient: false alarms, disconnected views, lost coverage, runtime and memory all belong in the comparison.

What the comparison must include

MapAnything already offers native multiview confidence. Doppelgangers++ and XDG address pairwise visual disambiguation and filter candidate relationships. These existing capabilities set the baseline: a combination of confidence and geometric signals needs to demonstrate additional value under the same inputs and false-removal constraints.

Evidence status · 2026-09-22

Ordinary engineering scenes completed: 1. Research scenes reviewed under protocol 0.2: 0. The engineering run used protocol 0.1; revision 0.2 preserves its engineering acceptance criteria. The execution record remains tied to its original version. Independent research validation is pending.

An independent reference

Correctness labels need evidence beyond the reconstruction being tested. A consistent cycle derived from the same global poses can repeat the same error. Lower optimization residuals alone cannot establish that visual ambiguity has been resolved.

Observe errors before deciding to repair

Initial geometry-error observations come before the gate for controlled repair. The requirements for multiple independent scenes, a second model and held-out improvement determine whether to attempt repair or expand the study. They do not require failures to be established before observation can begin.

Calibrate pose confidence separately

High-confidence pose errors need a separate calibration using motion-capture reference poses, with thresholds fixed for each model, checkpoint, preprocessing and four-view context. Surface-identity labels cannot replace pose-correctness labels. If calibration is insufficient, ordinary reference-confirmed errors remain reportable while the high-confidence classification stays unavailable.

Measure repair in the output

A repair must reduce independently confirmed errors in the final reconstruction while preserving views and geometric coverage. Removing candidate edges is an intermediate measurement. If the reconstruction backend already rejected an edge and its removal produces no independently verified output improvement, the outcome is no repair benefit.

A reproducible record

Each result should identify the scene and input provenance, model and code revisions, configuration, actual sample count, reference construction, measured costs and replay command. Failed or unresolved cases remain part of the record.

Start small, keep the reference separate

The first reference-backed control is a four-frame TUM RGB-D sequence with motion-capture poses. Pairwise ambiguity evaluation will use published surface-identity labels, with calibration and held-out data separated by physical scene. These are planned inputs; selected images, hashes and actual counts must be frozen before evaluation.

What the first run can establish

A successful ordinary-scene run establishes local execution and inspectable output. Evidence for a new diagnostic method requires comparison across independent scenes, strong alternatives and held-out data.

These upstream projects are starting points for the reproduction and method comparison. Their results remain attributable to their authors.

Earlier evidence

The First Generation report retains the published results, negative evidence and reproduction paths that inform this approach to evaluation.

What these numbers establish

The independent CPU review checked frame selection, reference matching, coordinate conventions and the six pair calculations against the saved model output. All pairs share cameras within one ordinary window. Reference noise has not been independently bounded; this evidence supports a recorded numerical comparison, with diagnosis, confidence calibration and repair still unvalidated.

Recompute without a GPU

The public CPU entry point rebuilds the reference rotations from recorded quaternions and recalculates the comparison from saved camera matrices. It needs Python and NumPy, with no model or image download. This checks the published record; running the model afresh uses the separate reproduction guide and external artifacts.

Why the assessment changed

GeoCond combines native confidence with geometry-related features extracted from backbone outputs, uses pose-graph cycle residuals as one supervision signal, and evaluates uncertainty and selective refinement. These model-derived signals are not independent ground-truth references, and its cross-backbone results do not establish universal improvement. The overlap in the core research question is nevertheless enough to withdraw our earlier claim of an unoccupied contribution.

What remains useful

GeometryAudit preserves the recorded runs, original protocols, compact camera records and CPU pose checks for inspection and reproduction. The R1 and TUM results retain their original scope; no new reliability or repair result is claimed. Any future research candidate will need its own question and prior-art assessment.