2012 11-16 image-videoanalysis_lecture08

1. 8: [email protected] Labs Russia16 2012, Computer Science Center

2. ? () , (image classification) , , , (object detection, localization) , (object segmentation)2 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: M. Everingham

10. ? - v.s. - Bottom-up v.s. top-down Pixel-based v.s. region-based/area-based Local v.s. global Feature-space based v.s. image-domain based Region-based v.s. edge-based ? - - 10 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

11. ? - v.s. - Bottom-up v.s. top-down Pixel-based v.s. region-based/area-based Local v.s. global Feature-space based v.s. image-domain based Region-based v.s. edge-based ? v.s. Automated v.s. user-directed11 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

12. 255192, 9603 255192, 8 D. Comaniciu, P. Meer: Robust Analysis of Feature Spaces: Color Image Segmentation, CVPR9712 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

13. - 4 Vincent Levesque, Texture Segmentation Using Gabor Filters http://www.cs.huji.ac.il/~simonp/papers/ip_project.pdf13 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

14. Thomas Hofmann, Jan Puzicha, Joachim M. Buhmann, Deterministic annealing for unsupervised texture segmentation, EMMCVPR 97, pp. 213-22814 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

16. , (motion) : , http://www.svcl.ucsd.edu/projects/motiondytex/demo.htm http://vcla.stat.ucla.edu/old/Barbu_Research/Motion_estim/index.html16 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

17. (depth) M. Domnguez-Morales, A. Jimnez-Fernndez, R. Paz-Vicente, A. Linares-Barranco and G. Jimnez-Moreno Stereo Matching: From the Basis to Neuromorphic Engineering17 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

23. k- G G R R : min23 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: http://www.cs.washington.edu/education/courses/cse576/08sp/lectures/segment.pdf

24. k- (k-means)1. k2. k (, )3. : 4. , 5. 4, 3 Java : http://home.dei.polimi.it/matteucc/Clustering/tutorial_html/AppletKM.html24 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: http://www.cs.washington.edu/education/courses/cse576/08sp/lectures/segment.pdf

25. k-: 1 Distance Metric: Euclidean Distance 5 4 k1 3 k2 2 1 k3 0 0 1 2 3 4 525 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: Lihi Zelnik-Manor

26. k-: 2 Distance Metric: Euclidean Distance 5 4 k1 3 k2 2 1 k3 0 0 1 2 3 4 526 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: Lihi Zelnik-Manor

27. k-: 3 Distance Metric: Euclidean Distance 5 4 k1 3 2 k3 k2 1 0 0 1 2 3 4 527 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: Lihi Zelnik-Manor

28. k-: 4 Distance Metric: Euclidean Distance 5 4 k1 3 2 k3 k2 1 0 0 1 2 3 4 528 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: Lihi Zelnik-Manor

29. k-: 5 Distance Metric: Euclidean Distance expression in condition 2 5 4 k1 3 2 k2 k3 1 0 0 1 2 3 4 5 expression in condition 129 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: Lihi Zelnik-Manor

30. k- K-means clustering based on intensity or color is essentially vector quantization of the image attributes Clusters dont have to be spatially coherent Image Intensity-based clusters Color-based clusters30 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: Lihi Zelnik-Manor

31. Distance based on color and position Source: K. Grauman31 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: Lihi Zelnik-Manor

32. k- Clustering based on (r,g,b,x,y) values enforces more spatial coherence32 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: Lihi Zelnik-Manor

33. k-Means: Converges to a local minimum of the error function (: K-means++) Memory-intensive Need to pick K Sensitive to initialization Sensitive to outliers Only finds spherical clusters33 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

34. Mean-shift for image segmentationUseful to take into account spatial information instead of (R, G, B), run in (R, G, B, x, y) space D. Comaniciu, P. Meer, Mean shift analysis and applications, 7th International Conference on Computer Vision, Kerkyra, Greece, September 1999, 1 1203. 197- http://www.caip.rutgers.edu/riul/research/papers/pdf/spatmsft.pdf More Examples: http://www.caip.rutgers.edu/~comanici/segm_images.html34 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: http://www.cs.washington.edu/education/courses/cse576/08sp/lectures/segment.pdf

35. Mean shift algorithmThe mean shift algorithm seeks modes or local maxima of density in the feature space Feature space image (L*u*v* color values)35 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

36. Mean shift clustering/segmentation Find features (color, gradients, texture, etc) Initialize windows at individual feature points Perform mean shift for each window until convergence Merge windows that end up near the same peak or mode36 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

37. Mean shift Search window Center of mass Mean Shift vector 37 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.Slide by Y. Ukrainitz & B. Sarel

43. Mean shift Search window Center of mass 43 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.Slide by Y. Ukrainitz & B. Sarel

44. Mean shift clustering Cluster: all data points in the attraction basin of a mode Attraction basin: the region for which all trajectories lead to the same mode 44 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.Slide by Y. Ukrainitz & B. Sarel

45. Mean shift segmentation results http://www.caip.rutgers.edu/~comanici/MSPAMI/msPamiResults.html45 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

47. Mean shift: Does not assume spherical clusters Just a single parameter (window size) Finds variable number of modes Robust to outliers Output depends on window size Computationally expensive Does not scale well with dimension of feature space47 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

48. Probabilistic clusteringBasic questions whats the probability that a point x is in cluster m? whats the shape of each cluster?K-means doesnt answer these questionsBasic idea instead of treating the data as a bunch of points, assume that they are all generated by sampling a continuous function This function is called a generative model defined by a vector of parameters 48 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: http://www.cs.washington.edu/education/courses/cse576/08sp/lectures/segment.pdf

49. Expectation maximization (EM) Goal find blob parameters that maximize the likelihood function: Approach: 1. E step: given current guess of blobs, compute ownership of each point 2. M step: given ownership probabilities, update blobs to maximize likelihood function 3. repeat until convergence EM demo: http://lcn.epfl.ch/tutorial/english/gaussian/html/index.html49 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: http://www.cs.washington.edu/education/courses/cse576/08sp/lectures/segment.pdf

53. Automatic graph cut q Cpq c p Fully-connected graph node for every pixel link between every pair of pixels, p,q cost cpq for each link cpq measures similarity o similarity is inversely proportional to difference in color and position53 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: http://www.cs.washington.edu/education/courses/cse576/08sp/lectures/segment.pdf

54. Segmentation by Graph Cuts w A B C Break Graph into Segments Delete links that cross between segments Easiest to break links that have low cost (similarity) similar pixels should be in the same segments dissimilar pixels should be in different segments54 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: http://www.cs.washington.edu/education/courses/cse576/08sp/lectures/segment.pdf

55. Min cut B A Link Cut set of links whose removal makes a graph disconnected cost of a cut: Find minimum cut gives you a segmentation55 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: http://www.cs.washington.edu/education/courses/cse576/08sp/lectures/segment.pdf

56. But min cut is not always the best cut... Cuts with lesser weight than the ideal cut Ideal Cut and it is NP-complete56 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

57. Normalized Cut B A Normalized Cut a cut penalizes large segments fix by normalizing for size of segments volume(A) = sum of costs of all edges that touch A57 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. Slide credit: http://www.cs.washington.edu/education/courses/cse576/08sp/lectures/segment.pdf

64. - Deformable contours/active contours/snakes , , Lei He, Zhigang Peng, Bryan Everding, Xun Wang, Chia Y. Han, Kenneth L. Weiss, William G. Wee, A comparative study of deformable contour methods on medical image segmentation, Image and Vision Computing 26 (2008) 14116364 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

73. Berkeley Segmentation DataSet [BSDS]D. Martin, C. Fowlkes, D. Tal, J. Malik. "A Database of Human Segmented Natural Images and its Application toEvaluating Segmentation Algorithms and Measuring Ecological Statistics", ICCV, 200173 Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

2012 11-16 image-videoanalysis_lecture08

Documents