Fast Robust Large-scale Mapping from Video and Internet Photo ...
Fast Robust Large-scale Mapping from Video and Internet Photo ...
Fast Robust Large-scale Mapping from Video and Internet Photo ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Keywords:<br />
1. Introduction<br />
The fully automatic modeling of large-<strong>scale</strong> environments has been a research<br />
goal in photogrammetry <strong>and</strong> computer vision since a long time. Detailed<br />
3D models automatically acquired <strong>from</strong> the real world have many uses<br />
including civil <strong>and</strong> military planning, mapping, virtual tourism, games, <strong>and</strong><br />
movies. In this paper we present a system approaching the fully automatic<br />
modeling of large-<strong>scale</strong> environments, either <strong>from</strong> video or <strong>from</strong> photo collections.<br />
Our system has been designed for efficiency <strong>and</strong> scalability. GPU implementations<br />
for SIFT feature matching, KLT feature tracking, gist feature<br />
extraction <strong>and</strong> multi-view stereo allow our system to rapidly process large<br />
amounts of video <strong>and</strong> photographs. For large photo collections, iconic image<br />
clustering allows the dataset to be broken into small closely related parts.<br />
The parts are then processed <strong>and</strong> merged, allowing tens of thous<strong>and</strong>s of photos<br />
to be registered. Similarly, for videos, loop detection finds intersections in<br />
the camera path, which reduces drift for long sequences <strong>and</strong> allows multiple<br />
videos to be registered.<br />
Recently mapping systems like Microsoft Bing Maps <strong>and</strong> Google Earth<br />
have started to use 3D models of cities in their visualizations. Currently these<br />
systems still require a human in the loop for delivering models of reasonable<br />
quality. Nevertheless they already achieve impressive results, modeling large<br />
areas with regular updates. However, these models are very low complexity,<br />
<strong>and</strong> do not provide enough detail for ground-level viewing. Furthermore the<br />
2