On-line Spectral Learning in Exploring 3D Large Scale Geo-Referred Scenes

Personalized navigation of 3D large scale geo-referred scenes has a tremendous impact in digital cultural heritage. This is a result of the recent progress in digitization technology which leads to the creation of massive digital geographic libraries. However, an efficient personalized 3D geo-referred architecture requires intelligent and on-line learning strategies able to dynamically capture user’s preferences dynamics. In this paper, we propose an adaptive spectral learning framework towards 3D navigation of geo-referred scenes. Spectral clustering presents advantages compared to traditional centerbased partitioning methods, such as the k-means; it effectively categorize nonGaussian, complex distributions, present invariability to shapes and densities and it does not depend on the similarity metric used since learning is performed through similarity matrices by exploiting pair-wise comparisons. The main difficulty, however, in incorporating spectral learning in a 3D navigation architecture is its static implementation. To handle this difficulty, we propose in this paper an adaptive framework through the use of adaptive spectral learning which tailors 3D navigation to user’s current needs.