View Clustering of Wide-Baseline N-views for Photo Tourism

The problem of view clustering is concerned with finding connected sets of overlapping views in a collection of photographs. The view clusters can be used to organize a photo collection, traverse through a collection, or for 3D structure estimation. For large datasets, geometric matching of all image pairs via pose estimation to decide on content overlap is not viable. The problem becomes even more acute if the views in the collection are separated by wide baselines, i.e. we do not have a dense view sampling of the 3D scene that leads to increase in computational cost of epipolar geometry estimation and matching. We propose an efficient algorithm for clustering of such many weakly overlapping views, based on opportunistic use of epipolar geometry estimation for only a limited number of image pairs. We cast the problem of view clustering as finding a tree structure graph over the views, whose weighted links denote likelihood of view overlap. The optimization is done in an iterative fashion starting from an minimum spanning tree based on photometric distances between image pairs. At each iteration step, we rule out edges with low confidence of overlap between the respective views, based on epipolar geometry estimates. The minimum spanning tree is recomputed and the process is repeated until there is no further change in the link structure. We show results on the images in the 2010 Nokia Grand Challenge Dataset that contains images with low overlap with each other.