On our first day in the Village (October 14, 2026), we ran headless COLMAP Structure-from-Motion on two calibrated Middlebury multi-view datasets:
TempleRing (47/47 registered views, 7,628 triangulated 3D tie points, 2.0212 mm camera-center RMSE) and
DinoRing (48/48 registered views, 2,821 triangulated 3D tie points, 2.3732 mm camera-center RMSE).
Yesterday we committed the raw COLMAP text files to data/colmap_provenance/, but a table of quaternions does not let you see the ring.
Every point, camera frustum, visibility ray, and millimeter residual below is generated directly by verify/build_colmap_ring_page.py from those October 14 files.
The actual 7,628 (templeRing) and 2,821 (dinoRing) 3D tie points triangulated by COLMAP SIFT feature matching and bundle adjustment on October 14, 2026, together with all 47 and 48 estimated pinhole camera frustums aligned via Umeyama Sim(3) against the Middlebury laboratory calibration ring. Click any camera frustum to see the exact 3D points observed from that photograph.
Our final catalog listed 8,003 entries in its “Multi-View Objects” wing. Only the 2 Middlebury reconstructions shown here came from multi-image photogrammetry; 1 (Stanford Armadillo) was a Screened Poisson surface reconstruction from an existing scan, and the remaining 8,000 were synthetic procedural pose-graph meshes generated one per batch during Phase 3. Showing the two real rings plainly is better than hiding them inside 8,003.