| <Poster Width="1734" Height="1041"> |
| <Panel left="3" right="125" width="563" height="469"> |
| <Text>Introduction</Text> |
| <Text>IGoal: Recognize the an image’s location by matching to a database</Text> |
| <Text>IChallenges: matching is time consuming; image retrieval is noisy</Text> |
| <Text>IPrevious Approaches: image retrieval based & direct matching</Text> |
| <Text>IOur Approach:</Text> |
| <Text>I Use an image graph to learn local similarity functions</Text> |
| <Text>I Encourage diversity in top ranked results</Text> |
| <Figure left="95" right="302" width="386" height="280" no="1" OriWidth="0.351211" OriHeight="0.194296 |
| " /> |
| </Panel> |
|
|
| <Panel left="2" right="601" width="561" height="398"> |
| <Text>Image Graphs</Text> |
| <Text>INodes are images</Text> |
| <Text>IOnly geometrically</Text> |
| <Text>consistent images</Text> |
| <Text>are connected</Text> |
| <Text>IEdge weights defined</Text> |
| <Text>by Jaccard Index</Text> |
| <Text>N(a,b)J(a,b): </Text> |
| <Text>N(a)+N(b)−N(a,b),</Text> |
| <Text>and thresholded to</Text> |
| <Text>improve robustness</Text> |
| <Text>IOn the right: an</Text> |
| <Text>example image graph</Text> |
| <Text>on Dubrovnik dataset</Text> |
| <Text>(red nodes are center</Text> |
| <Text>images selected)</Text> |
| <Figure left="197" right="629" width="352" height="362" no="2" OriWidth="0.351788" OriHeight="0.287879 |
| " /> |
| </Panel> |
|
|
| <Panel left="572" right="125" width="578" height="229"> |
| <Text>Overview of Approach</Text> |
| <Text>ITraining:</Text> |
| <Text>I 1. Compute a covering of the graph with a set of subgraphs (select center images</Text> |
| <Text>or neighborhoods in the image graph).</Text> |
| <Text>I2. Learn and calibrate an SVM-based distance metric for each subgraph.</Text> |
| <Text>ITesting:</Text> |
| <Text>I 3. Use the models in step 2 to compute the distance from a query image to each</Text> |
| <Text>database image, and generate a ranked shortlist of possible image matches.</Text> |
| <Text>4. Perform geometric verification sequentially with the top database images in the</Text> |
| <Text>shortlist.I</Text> |
| </Panel> |
|
|
| <Panel left="572" right="361" width="578" height="635"> |
| <Text>Generating Ranking Results</Text> |
| <Text>IRanked neighborhoods are concatenated to form a ranking list of all DB images</Text> |
| <Text>IOrder within each neighborhood determined by BoW similarity</Text> |
| <Text>IGoal: to have the first true match appear in ranked shortlist as early as possible.</Text> |
| <Text>IComparison of BoW image retrieval ranking and our learned ranking:</Text> |
| <Figure left="600" right="493" width="516" height="269" no="3" OriWidth="0.387543" OriHeight="0.154635 |
| " /> |
| <Text>IRanking can be further improved by enforcing diversity in top results: pick the next</Text> |
| <Text>image conditioned on previous one failing to match</Text> |
| <Figure left="601" right="820" width="520" height="158" no="4" OriWidth="0.374856" OriHeight="0.0882353 |
| " /> |
| </Panel> |
|
|
| <Panel left="1158" right="125" width="557" height="763"> |
| <Text>Experiments</Text> |
| <Figure left="1233" right="160" width="410" height="71" no="5" OriWidth="0.378316" OriHeight="0.0432264 |
| " /> |
| </Panel> |
|
|
| <Panel left="1159" right="891" width="561" height="112"> |
| <Text>Reference</Text> |
| <Text>[1] Y. Li, N. Snavely, and D. Huttenlocher. Location recognition using prioritized</Text> |
| <Text>feature matching. In ECCV, 2010.</Text> |
| <Text>[2] T. Sattler, T. Weyand, B. Leibe, and L. Kobbelt. Image retrieval for image-based</Text> |
| <Text>localization revisited. In BMVC, 2012.</Text> |
| </Panel> |
|
|
| </Poster> |
|
|
