3D Reconstruction of Buildings From Images with Automatic Facade Refinement
Author
Larsen, Christian Lindequist
Term
4. term
Publication year
2010
Submitted on
2010-06-03
Pages
125
Abstract
Dette projekt undersøger, hvordan man kan skabe 3D-modeller af virkelige bygninger ud fra fotos for at gøre interaktive visualiseringer af boliger mere praktiske, fx på websider. For at dette kan lade sig gøre, skal rekonstruktionen gøres enklere og den manuelle indsats mindskes. Vi har udviklet et proof-of-concept-system, der tager et uordnet sæt billeder af en bygning, beregner bygningens 3D-form og kamerabevægelse (ofte kaldet struktur-fra-bevægelse), og lader en bruger hurtigt samle en grov, tekstureret 3D-model. Det primære bidrag er en ny metode, der automatisk tilføjer facadedetaljer—som indrykkede vinduer og døre—til den grove model. Den finder disse elementer ved at analysere facadens udseende med metoder fra billedbehandling og mønsterklassifikation. I tests på flere bygninger producerede systemet forfinede 3D-modeller, og den automatiske facademetode detekterede og rekonstruerede i gennemsnit korrekt cirka 89% af de indrykkede vinduer. Noget brugertilpasning er stadig nødvendig—fx skal dybden af de indrykkede områder angives af brugeren—men samlet set mindsker metoden manuelt arbejde sammenlignet med eksisterende brugerassisterede metoder.
This project explores creating 3D models of real buildings from photos to make interactive visualizations of properties—such as on websites—more practical. To make this feasible, reconstruction needs to be simpler with less manual effort. We built a proof-of-concept system that takes an unordered set of images of a building, estimates the building’s 3D shape and camera motion (a step often called structure-from-motion), and lets a user quickly assemble a coarse, textured 3D model. The main contribution is a new method that automatically adds facade details—such as recessed windows and doors—to that coarse model. It identifies these features by analyzing the facade’s appearance using image processing and pattern classification techniques. In tests on several buildings, the system produced refined 3D models, and the automatic facade method correctly detected and reconstructed about 89% of recessed windows on average. Some user input is still required—for example, the depth of recessed regions must be specified—but overall the method reduces manual work compared with existing user-assisted approaches.
[This abstract was generated with the help of AI]
Keywords
Documents