3D Reconstruction from Catadioptric Camera [ OMNIVIS'05, CVPR'07, CVIU'08, CVPR'08 ]

Email: Maxime { dot } Lhuillier { at } free { dot } fr
Address: Institut Pascal, UMR 6602 CNRS/UBP/IFMA,
Campus Universitaire des Cézeaux,
63171 Aubière Cedex - France.
Tel: +33 (0)4 73 40 75 93
Fax: +33 (0)4 73 40 72 62

Last Modification on June 30th 2008.
Go back to my HOME PAGE.

About it

First, still images are acquired by a roughly equiangular camera, alternating a step forward and a shot. Second, the sequence geometry is estimated with a central camera model. The geometry includes all camera poses, some intrinsic parameters and a sparse set of 3D points of the scene. Third, many local 3D models along the sequence are reconstructed with 3 consecutive images. Catadioptric images are reprojected on a virtual cube and (quasi) dense stereo is applied between cube faces. At the end of this step, the scene part visible from the 3 images of a local model is approximated by a list of textured and planar patches (patches are small squares in the catadioptric image and are planar parts in 3D). Last, a global 3D model of the scene is obtained by an adequate selection on all patches of all local models.
The CVPR'07 paper describes the overall method, while the CVIU'08 paper focuses on the geometry estimation step (for both central and non-central camera models).

The following results are obtained by the catadioptric camera below (the 0-360 mirror with the Nikon Coolpix 8700). This camera is hand-held and mounted on a monopod.


Clic on the following icons to visualize a mpg/mp4 video.
The videos shows a walkthrough in the global 3D model of the scene (in the neighborhood of the original view points). Every camera pose is represented by a cone with gray-levels. The window at the bottom left corner shows a local and top view of the geometry estimation (reference camera locations are blue squares and 3D points are black points).
There are also videos which show the catadioptric images (input) of the algorithm (panoramic image are also given on the top to help understand the scene, but they are not used by the method).

Old Town Sequence: 354 1632x1224 images (input)

Small Room Sequence: 18 1632x1224 images

Also there is an other video obtained with an alternative method using a generic camera model (CVPR'08).

Church Sequence: 208 1632*1224 images (input)