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.
Figure 4: Left: view of the 3D model <strong>and</strong> the 9025 registered cameras. Right: model <strong>from</strong><br />
the interior of Ellis Isl<strong>and</strong> erroneously labeled as “Statue of Liberty”.<br />
extending the registration to all cluster images in the clusters corresponding<br />
to the registered iconics. This process takes advantage of feature matches<br />
between the non-iconic images <strong>and</strong> their respective iconics. This process<br />
again uses the 2D matches to the iconic to determine the 2D-3D correspondences<br />
for the image. It then applies ARRSAC (Section 4.1.2 to determine<br />
the camera s registration. Given the incremental image registration during<br />
this process drift inherently accumulates. Hence we periodically apply<br />
bundle adjustment to perform a global error correction. In our current system<br />
bundle adjustment is performed after 25 images have been added to a<br />
sub-model. This saves significant computational resources while in practice<br />
keeping the accumulated error at a reasonable level. Detailed results of our<br />
3D reconstruction algorithm are shown in Figures 4-5.<br />
5. Dense Geometry<br />
The above described methods for video <strong>and</strong> for photo collections provide<br />
a registration for the camera poses of all video frames or photographs.<br />
These external <strong>and</strong> internal camera calibrations are the output of the sparse<br />
28