1 / 35

View Selection

View Selection. Presented by Marlene Shehadeh . Advanced Topics in Computer Vision ( 048921 ) . Winter 2011-2012. Problem. Goals . arrange: Icon BCS . A Review of: Discovering Favorite Views of Popular Places with Iconoid Shift Tobias Weyand and Bastian Leibe ICCV 2011.

yelena
Download Presentation

View Selection

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. View Selection Presented by Marlene Shehadeh Advanced Topics in Computer Vision (048921) Winter 2011-2012

  2. Problem

  3. Goals • arrange: • Icon • BCS

  4. A Review of:Discovering Favorite Views of Popular Places with Iconoid Shift Tobias Weyand and Bastian LeibeICCV 2011

  5. Iconoid shift algorithm outline

  6. Medoid shift • For a given kernel and distance function • Center is always a point in the set • The clustering in performed iteratively

  7. Homography Overlap Distance • Modes search for image with maximal overlap with its neighbors • The distance measure must: • Reward similar views • Penalize different view ( zoom, pan) • Consider the overlap region • are the areas of images and bounding boxes around the inliers

  8. Homeography distance

  9. Transitive Homography Overlap Distance • Medoid shift requires the distance of each pair of images in the set • Direct feature matching distance calculation is very expensive • Solution: • Represent local neighborhood by a tree • Compute distance along edges • Calculate the distance using the tree path

  10. Hinge kernel • The kernel : • Cuts off images with distance greater than the threshold

  11. Implementation efficiency • Local Exploration and Minimum Spanning Tree Construction • node i stores overlap region with the root • homography overlap distances to root are computed by propagating the overlap region • only O(N) propagation steps have to be performed • Homography Overlap Propagation (HOP) • For parent i the homography overlap is propagated to all nodes j in its subtree • the transitive propagation scheme is used to compute the distances between all nodes n and m that have common parent i • O(N) memory complexity, O(N2) time complexity.

  12. Experimental results • Dataset of 500k images of paris from tourist pictures • Initial seed set is determined by Min-Hash Seed Generation

  13. results

  14. Kernel effect

  15. EVALUATION • Experimental evaluation • Many visual results • Comparison to existing methods • Subsystem tests • Large random dataset Novel approach in image soft clustering Novel combination of existing parts Well written Not self contained Technically convincing

  16. A Review of:Selecting Canonical Views for View-Based 3-D Object Recognition T. Denton et alICPR 2004

  17. algorithm outline • Given a set of views P and a similarity function S • Construct a graph: • Views are vertices • Edges are proportional to nodes similarity • Find bounded canonical subset(BCS) • Minimize the sum of edges within BCS • Maximize the sum of edges between BCS and the rest of the set

  18. Problem of maximizing weight edges is NP hard • Approximate Solution: • Semidefinite programming (SDP) • Normalized cut

  19. Normalized cut • patter is assigned an indicator • if the pattern belongs to the BCS • Cut edge maximization • Intra edge minimization

  20. Reformulation as quadratic problem • SDP is used to solve this problem

  21. Similarity measure

  22. Experimantal results • 2D images were acquired from 3D synthetic objects • Each object has 19 views acquired by sampeling the view sphere • The resulting BCS views were compared to the rest of the set and ranked. • In 90.6% of the cases the correct canonical view was among the top 6 ranks.

  23. when different objects share similar views the correct canonical view may not be top ranked • If the bounds are set too low for a complex object, whole classes of object views are not represented in the object’s BCS

  24. Future work • Evaluate the method quantitivly • Study the effect of set size and boundaries on performance • Adjust the algorithm to use a simpler matching method to replace the many to many complex method used

  25. EVALUATION • Experimental evaluation • Synthetic results only • No comparison to other methods • No quantative results Extention to previous works for summarizing sets Novel combination of existing parts Not self contained

  26. A Review of:Finding Iconic Images Tamara L. Berg Alexander C. Berg CVPR’09Internet Vision Workshop, 2009

  27. outline

  28. Ranking • Learn model • The image consists of a rectangular foreground, and background • Possible layouts are examined • High score is give to an image with icon layout • Top ranked images are are used to calculate similarity

  29. Experimental results • learning set of 500 images • 17 categories of 100k initial images

  30. References • T. Weyand and B. Leibe ,”Discovering Favorite Views of Popular Places with Iconoid Shift”, ICCV 2011. • T. Denton, M. Demirci, J. Abrahamson, A. Shokoufandeh,and S. Dickinson. “Selecting Canonical Views for View-based 3D Object Recognition”. In ICPR, 2004. • T. Berg and A. B. Berg. “Finding Iconic Images”. In CVPR’09,Internet Vision Workshop, 2009. • O. Chum and J. Matas. “Large-scale discovery of spatiallyrelated images”. In PAMI, 2010.

More Related