One research stream in the camera project is investigating the automatic generation of CAD models from video streams. A notorious problem in stereo reconstruction is calibration. With calibrated cameras it is possible to provide metric reconstruction of scenes. With uncalibrated or semi-calibrated cameras it is possible to reconstruct scenes up to a projective transformation. The research at KTH has focussed on methods for estimation of the fundamental matrix that specifies the epipolar geometry between cameras or the position of the camera between subsequent frames. Most stereo approaches use epipolar geometry to estimate the disparity field. Through use of differential geometry it is possible to provide a unified framework that enable integration of stereo and motion information. Recent work has investigated how differential geometry might be used for scene recovery.
Reconstruction of already built environments based on sparse or dense image sequences is practical if it is sufficiently automated (user interactions are reduced to minimum). IGD's use of existing reconstruction systems [CaReSs] shows that registering multiple images is the most time consuming operation at the present stage and it is required to be automated first. (The work has so far concentrated on this particular step). To initialize the reconstruction process a set of reliable matches between images need to be provided. At this first step no registration or calibration information is available and so the robustness of matches is the most important issue. However, images of objects that contain weakly textured areas or smooth edges do not contain sufficient amounts of such discriminative features.
IGD has defined a new point representation based on local chromatic gradients. This point representation is stable in every direction, since the same amount of information is integrated from every radius. Also, it integrates geometric and chromatic information in the point neighbourhood thus creating a robust semi-local descriptor.
In the figure below two different views of the Bornholm church are presented with matched points. Points were selected in areas with low chromatic variation and thus show the ability of the algorithm to cope with such points.
