Explore historic architecture, align photographs and navigate a real 3D reconstruction.
WHO IT HELPS
Heritage, inspection and vision teams comparing viewpoints or processing photographs into 3D.
WHAT IT DOES
Explore historic architecture, align photographs and navigate a real 3D reconstruction.
DEMO OUTPUT
ES/EN demo with 104 attributed photographs, fresh matching, a navigable point cloud, PLY/JSON, report and presentation. The local application also reconstructs new batches.
Actual recording · Synthetic English narration and captions · Attributed public photographs
THE PROBLEM
Building photographs taken from different viewpoints and periods cannot be compared directly. Geometry, appearance differences and reliable evidence need to be separated.
THE WORKFLOW
Choose El Pilar, Notre-Dame or the Alhambra; compute new ORB correspondences in the browser, blend aligned views and review candidate differences. Also explore a real COLMAP reconstruction of El Pilar.
01Attributed or own photographs
02Fresh matching and alignment
03Visual review and 3D point cloud
04Evidence export
RECORDED EVIDENCE
Results, with context.
Geometric registration on public photographs
Check
Cases
Registered cameras / modern photos
40 of 65
Python pairs with recorded alignment
3 of 3
Dataset
heritage-atlas-v1
Completed cases
3/3
Median / p95 per case
0.236 s / 0.244 s
Evaluation record (UTC)
2026-10-10
Actual CPU execution on attributed Wikimedia Commons photographs. Median/p95 covers three complete Python SIFT analyses; recorded reconstruction time is 153.34 s. The interactive demo separately times each new ORB computation. Reprojection error measures internal fit, without metric ground truth or independent change annotations.
40 of 65 modern photographs registered in the main model: 1,680 sparse points and 0.429 px median reprojection error. Arbitrary scale, without a mesh or metric ground truth. Differences need review; historical localisation was rejected as ambiguous.
Recorded demo
Recorded demo