Introduction IGoal: Recognize the an image’s location by matching to a database IChallenges: matching is time consuming; image retrieval is noisy IPrevious Approaches: image retrieval based & direct matching IOur Approach: I Use an image graph to learn local similarity functions I Encourage diversity in top ranked results

Image Graphs INodes are images IOnly geometrically consistent images are connected IEdge weights defined by Jaccard Index N(a,b)J(a,b): N(a)+N(b)−N(a,b), and thresholded to improve robustness IOn the right: an example image graph on Dubrovnik dataset (red nodes are center images selected)

Overview of Approach ITraining: I 1. Compute a covering of the graph with a set of subgraphs (select center images or neighborhoods in the image graph). I2. Learn and calibrate an SVM-based distance metric for each subgraph. ITesting: I 3. Use the models in step 2 to compute the distance from a query image to each database image, and generate a ranked shortlist of possible image matches. 4. Perform geometric verification sequentially with the top database images in the shortlist.I Generating Ranking Results IRanked neighborhoods are concatenated to form a ranking list of all DB images IOrder within each neighborhood determined by BoW similarity IGoal: to have the first true match appear in ranked shortlist as early as possible. IComparison of BoW image retrieval ranking and our learned ranking:

IRanking can be further improved by enforcing diversity in top results: pick the next image conditioned on previous one failing to match

Experiments

Reference [1] Y. Li, N. Snavely, and D. Huttenlocher. Location recognition using prioritized feature matching. In ECCV, 2010. [2] T. Sattler, T. Weyand, B. Leibe, and L. Kobbelt. Image retrieval for image-based localization revisited. In BMVC, 2012.