Presentation is loading. Please wait.

Presentation is loading. Please wait.

Normalized Cuts Demo Original Implementation from: Jianbo Shi Jitendra Malik Presented by: Joseph Djugash.

Similar presentations


Presentation on theme: "Normalized Cuts Demo Original Implementation from: Jianbo Shi Jitendra Malik Presented by: Joseph Djugash."— Presentation transcript:

1 Normalized Cuts Demo Original Implementation from: Jianbo Shi Jitendra Malik Presented by: Joseph Djugash

2 Outline Clustering Point  The Eigenvectors  The Affinity Matrix  Comparison with K-means Segmentation of Images  The Eigenvectors  Comparison with K-means

3 Clustering – How many groups are there? Out of the various possible partitions, which is the correct one?

4 Clustering – Why is it hard? Number of components/clusters? The structure of the components? Estimation or optimization problem?  Convergence to the globally correct solution?

5 Clustering – Example 1 Optimal? How do we arrive at this Clustering?

6 What does the Affinity Matrix Look Like?

7 The Eigenvectors and the Clusters Step-Function like behavior preferred! Makes Clustering Easier.

8 The Eigenvectors and the Clusters

9 Clustering – Example 2 Dense Square Cluster Sparse Square Cluster Sparse Circle Cluster

10 Normalized Cut Result

11 The Affinity Matrix

12 The Eigenvectors and the Clusters

13 K-means – Why not? e1 e2 Input Eigenvectors Affinity Matrix Eigenvector Projection NCut Output K-means Output K-means Clustering? Possible but not Investigated Here.

14 K-means Result – Example 1

15 K-means Result – Example 2

16 Varying the Number of Clusters k = 3k = 4k = 6 K-means N-Cut

17 Varying the Sigma Value σ = 3σ = 13σ = 25

18 Image Segmentation – Example 1 Affinity/Similarity matrix (W) based on Intervening Contours and Image Intensity

19 The Eigenvectors

20 Comparison with K-means Normalized Cuts K-means Segmentation

21 How many Segments?

22 Good Segmentation (k=6,8)

23 Bad Segmentation (k=5,6) Missing Edge Bad Edge Choice of # of Segments in Critical. But Hard to decide without prior knowledge.

24 Varying Sigma – (σ= Too Large)

25 Varying Sigma – (σ= Too Small) Choice of Sigma is important. Brute-force search is not Efficient. The choice is also specific to particular images.

26 Image Segmentation – Example 2

27 Normalized Cuts K-means Segmentation

28 Image Segmentation – Example 3

29 Normalized Cuts K-means Segmentation

30 Image Segmentation – Example 4

31 Normalized Cuts K-means Segmentation

32 Image Segmentation – Example 5

33 Normalized Cuts K-means Segmentation

34 Image Segmentation – Example 6

35 Comparison with K-means Normalized Cuts K-means Segmentation

36 The End…

37 The Eigenvectors and the Clusters Eigenvector #1 Eigenvector #2Eigenvector #3Eigenvector #4Eigenvector #5


Download ppt "Normalized Cuts Demo Original Implementation from: Jianbo Shi Jitendra Malik Presented by: Joseph Djugash."

Similar presentations


Ads by Google