Section: New Results

Omnidirectional vision

Calibration of medical endoscopes.

Some endoscopes give a very wide field of view, in order to fully inspect e.g. vessels in the human body. Often, endoscopes are used for visualization purposes only, less frequently for making measurements. Nevertheless, an approximate calibration is required even for visualization, in order to present the wide-field of view images, which include strong distortion, in a more common way and without these distortions. In [26] we propose a very pragmatic method for endoscope calibration, well suited for non-expert users, requiring a single image of a planar grid. This is joint work with João Barreto from Coimbra University. In another joint work, we developed a more general calibration approach, requiring multiple images of a planar grid but allowing to calibrate a wider range of camera models, namely any central catadioptric camera [36] . It is based on a recent result of ours, showing the existence of fundamental matrices and plane homographies, for any such camera (P. Sturm and J. Barreto, General Imaging Geometry for Central Catadioptric Cameras, ECCV'08)

Geolocalization using Skylines from Omni-Images.

In this joint work with MERL [43] , the goal is to provide an image-based method for geolocalization in cities, where the GPS is known to be unreliable. The developed method exploits the perhaps surprising fact that skylines, as seen from city streets, are rather characteristic for the camera location. We use upward-looking fisheye cameras with a roughly hemispheric field of view; in the acquired images, skylines are extracted relativelily reliable by segmenting the regions belonging to the sky. Nowadays, coarse 3D city models are easily available; given such a model, it is simple to generate the skyline that should be seen from any given location in the city. This is used in our geolocalization approach, where the camera location is found by comparing skylines extracted in images with those generated from candidate camera locations. This approach has been shown to be very robust and reasonable accurate, over image sequences corresponding to camera displacements of hundreds of meters.