Randomized Structure from Motion
Based on Atomic 3D Models from Camera Triplets

Michal Havlena, Akihiko Torii, Jan Knopp, and Tomas Pajdla

All videos are compressed using x264vfw codec: http://sourceforge.net/projects/x264vfw/
If you cannot play them, please try
VLC (win, mac & linux): http://www.videolan.org/vlc/
K-lite (win): http://www.codecguide.com/download_kl.htm

In case you do not have a VRML viewer and plugin, the one suitable for your environment can be found at: http://cic.nist.gov/vrml/vbdetect.html
Octaga Player (win, mac & linux): http://www.octaga.com/joomla/index.php?option=com_content&task=view&id=17&Itemid=40

'CASTLE data set'

The 3D models of several partial reconstructions (see Figures 7, 8, and 9) together with sample input images:

Open the VRML model

Right part of the St.Vitus Cathedral and other buildings in the square were reconstructed from 90 cameras, another 49 cameras were connected during gluing.

Open the VRML model

Part of the Old Town Square with the clock tower was reconstructed from 69 cameras, another 39 cameras were connected during gluing.

Open the VRML model

Entrance to the Prague Castle was reconstructed from 60 cameras, another 49 cameras were connected during gluing.

'DALIB data set'

The 3D model of the complete reconstruction (see Figure 3) together with sample input images:

Open the VRML model

Partial reconstruction containing all 39 cameras from selected atomic 3D models was extended with 25 missing cameras during gluing.

Videos generated from our VRML model:

Open the AVI video

Open the AVI video

Left: Sparse model using voxels located at reconstructed 3D points positions. Right: Dense model constructed using a multiview stereo reconstruction algorithm (similar to Y. Furukawa and J. Ponce. Accurate, dense, and robust multiview stereopsis. In CVPR 2007.) Transparent textures are the input images located according to the computed camera poses.

All 13 atomic 3D models selected by our technique as the seeds for merging: