Intelligent Database Systems Lab 國立雲林科技大學 National Yunlin University of Science and Technology VisualRank- Applying PageRank to Large-Scale Image Search Presenter : Chien-Hsing Chen Author: Yushi Jing Shumeet Baluja PAMI (IEEE Transactions on Pattern Analysis and Machine Intelligence)
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 2 Outline Motivation Objective Method Experiments Conclusion Comment
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 3 Motivation retrieved images may be not fitting (satisfactory) or not diverse The news shows a disappointed salesman of Coca Cola returns from his Middle East assignment. A friend asked, “Why weren’t you successful with the Arabs?” How the image could be retrieved ?
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 4 Example 2
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 5 Objective improve quality of image retrieval by rearrange the results of Google search engine incorrect retrieval d80 Coca Cola diversity (retrieved images should be different) You should know: 1. adjacency matrix, matrix product 2. eigenvector, PageRank()
Intelligent Database Systems Lab N.Y.U.S.T. I. M. Rearrange previous works Query-to-images in this paper Images-to-images 6 Main idea 1/
Intelligent Database Systems Lab N.Y.U.S.T. I. M. Rearrange Images-to-images 1. similar local features Web site source my homepage V.S. Yahoo 2. diversity 7 Main idea 2/
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 8 Images-to-images 1/ Top ranked images : Adjacency matrix How to connect between vertexes ? (how to build edge sets) x1 x2 x3 x4 x5 x1 x2x3 x4 x5
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 9 Images-to-images 2/ Top ranked images : Adjacency matrix How to give the scores between vertexes ? x1 x2 x3 x4 x5 x1 x2x3 x4 x5
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 10 How to connect between vertexes ? Common local features 1/2 which pair has most number of common (similar) local features? (a) local features, such as hands, eyes, are similar local features are very different (g)
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 11 Common local features 2/2
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 12 image relationship Which image has most number of common (similar) local features? A image of which features are similar to the features in the other images. The image is important × = [n × n] matrixeigenvector The entry is evaluated by “local features” uniform ?
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 13 PageRank PageRank() concerns the properties of “Hub” and “Authority” Web sites appearing in front of the Google responds are more important than that appearing in back of the ones. d
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 14 image diversity Top ranked images with respect to diversity:
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 15 Experiments
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 16 c
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 17 c
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 18 Conclusion arrange the images from the results of Google search engine
Intelligent Database Systems Lab N.Y.U.S.T. I. M. 19 Comment Advantage The aspect is novel and easy to implement. Drawback less discussion in diversity Application responds of search engine an option is to cluster the resulted images