Optimal Reduction of Large Image Databases for Location Recognition

For some computer vision tasks, such as location recognition on mobile devices or Structure from Motion (SfM) computation from Internet photo collections, one wants to reduce a large set of images to a compact, representative subset, sometimes called ``key frames'' or ``skeletal set''. We examine the problem of selecting a minimum set of such key frames from the point of view of discrete optimization, as the search for a minimum connected dominating set (CDS) of the graph of pair wise connections between the database images. Even the simple minimum dominating set (DS) problem is known to be NP-hard, and the constraint that the dominating set should be connected makes it even harder. We show how the minimum DS can nevertheless be solved to global optimality efficiently in practice, by formulating it as an integer linear program (ILP). Furthermore, we show how to upgrade the solution to a connected dominating set with a second ILP if necessary, although the complete method is no longer globally optimal. We also compare the proposed method to a previous greedy heuristic. Experiments with several image sets show that the greedy solution already performs remarkably well, and that the optimal solution achieves roughly 5% smaller key frame sets which perform equally well in location recognition and SfM tasks.