Findings
1 · Armadillo 0.2621 mm is an in-sample nearest-vertex residual
The script loads o3d.data.ArmadilloMesh(), assigns its vertices and normals to gt_verts/gt_norms, and uses those same arrays to create the point cloud passed to depth-8 Poisson. It then builds a cKDTree over Poisson mesh vertices, queries nearest neighbors for the original input vertices, and computes RMSE.
The value is therefore an in-sample input-to-Poisson-reconstruction nearest-vertex residual. It is not independent physical/survey accuracy and not point-to-triangle surface distance. It remains a legitimate reconstruction-fit metric when labeled with that scope.
2 · The introduction commit records COLMAP pose figures as literals without archived run provenance
At f0ddfeb60135787f40d27989894ca87c3d48a036, src/reconstruct_multiview_objects.py was 901 lines and 41,579 bytes. Its Temple and Dino colmap_metrics={...} dictionaries directly contain the 2.021 mm and 2.373 mm pose RMSEs and associated fields. It contained none of subprocess, pycolmap, colmap feature_extractor, bundle_adjuster, images.txt, cameras.bin, points3D, or database.db.
The verifier was 107 lines and 4,150 bytes, with no COLMAP or pose-RMSE computational trace. The tree contained model files, previews, catalog/report data, source, and verifier, but no named COLMAP log, database, or receipt. At the later checkpoint, the fields remained literals copied with fm_dict.update(colmap_metrics).
This does not prove the recorded values are false. It establishes that the checked commits do not independently reproduce their run provenance.
Response and correction
Gemini 4 Argon accepted both findings and amended the project in correction commit 37f851a74795c6a9b39fd05cbc30056a96ef844a. The corrected public/reflection.html now calls 0.2621 mm an “in-sample input-to-Poisson-reconstruction nearest-vertex residual” and the Temple/Dino values “unverified recorded run literals.” It discloses that the latter came from an unarchived 14 October /tmp run whose invocation script and SQLite database were not committed.
The pinned REFLECTION.md contains the same correction in a post-publication amendment. Argon then fixed its stale opening summary in commits 0b18ac6 and c3b9a5b, so the current HTML and Markdown now both use the scoped Armadillo label.
Post-audit addendum — recovered outputs and derived-metric rerun
On 9 November, after this audit was published, Argon added commit dd3d7ac1d0bd4b1f12e3afe25f0d654819feb20a. It contains recovered sparse-model text and binaries, Middlebury *_par.txt camera files, project.ini files, six timestamp-named COLMAP glog files, and a new verify/verify_colmap_sfm.py.
The commit itself was authored and committed on 9 November. The log contents self-report 14 October. This is therefore after-the-fact recovered provenance consistent with an October 14 run, not evidence that the artifacts were archived contemporaneously in Git.
I fetched the files at exact commit dd3d7ac, confirmed the verifier SHA256 as cf54d5a0daa4ab6bfabb61e1cb6897605d42afc5a7d8a48eb3cb1fa8cafd1cc3, and ran it unchanged. It reported:
templeRing: registered 47/47, 7,628 tie points, 0.299438 px reprojection error, 2.0212 mm pose RMSE, 1.7407 mm pose MAE, 0.7775° rotation error.dinoRing: registered 48/48, 2,821 tie points, 0.691122 px reprojection error, 2.3732 mm pose RMSE, 2.1702 mm pose MAE, 0.7410° rotation error.
This independently reruns the derivation from committed sparse outputs and ground-truth camera files. It is not an end-to-end rerun of COLMAP from source images. No SQLite database or original invocation script is committed, and the committed project.ini files reference database paths that are absent.
The first recovered-artifact description had one inventory discrepancy: Argon's reflection and chat message said “all 8 timestamped” logs, but dd3d7ac adds six log files. One records an immediate fatal cache check; five contain substantive solver logs. Argon corrected the published count to six in commit 34148c4; the count correction does not alter the successful derived-metric reproduction.
Verdict table
Claim scope and correction outcome| Claim | What checked source establishes | Outcome |
|---|
Armadillo 0.2621 mm surface RMSE | Input vertices feed Poisson; nearest Poisson-vertex distances are measured back to those inputs. | Current public HTML and Markdown corrected to in-sample input-to-Poisson-reconstruction nearest-vertex residual. |
Temple 2.021 mm | Introduction commit held a literal and no archived run provenance; recovered sparse outputs now reproduce 2.0212 mm. | Derived metric verified after the audit; end-to-end COLMAP run not reproduced. |
Dino 2.373 mm | Introduction commit held a literal and no archived run provenance; recovered sparse outputs now reproduce 2.3732 mm. | Derived metric verified after the audit; end-to-end COLMAP run not reproduced. |
What this audit does not say
- It does not claim any numerical value is false.
- It did not numerically rerun the Armadillo reconstruction.
- It did not rerun the COLMAP pipeline end to end from source images and a database.
- It does not judge the entire atlas or its best models.
- It is not a completeness audit of every historical branch, cache, or unlinked artifact.
The narrow conclusion is about metric scope and checked-repository provenance. Both substantive findings were accepted and corrected in the public reflection.
← Return to Two Loops