Classification

Question 1

Segmentation with classification

Answer 1

• classifiers estimate likelihood functions from samples
• classification with known number of classes and class boundaries (Thresholding)
• classification with known number of classes (Gaussian mixture models, k-means clustering)
• classification without prior knowledge (Mean-shift clustering, self-organizing maps)

Question 2

segmentation by classification

Answer 2

• 2-class problem: object class, background class

- optimal solution minimizes wrong assignments of voxels v to classes c0 and cb

Question 3

gaussian mixture models

Answer 3

• applied if intensity distribution is combination of different distributions

Question 4

Expectation Maximization

Answer 4

• assume K classes is known
• 1. E-step: compute estimates p_ij for intensity samples Ij belonging to class i
• 2. M-step: compute new parameters (mean, variance,, ) of classes
• local optimum, need to have good starting points for mean, variance,…

Question 5

K-Means Clustering

Answer 5

• optimal selection of cluster centers and assignment to classes (cluster centers recomputed after each sample assignment)
  + fast convergence
  + can be used as preprocessing step for segmentation
• strongly depends on init (trail and error)

Question 6

Mean Shift Clustering

Answer 6

• ## attempts to find all possible cluster centers in feature space

Question 7

Mean Shift Clustering

Answer 7

• attempts to find all possible cluster centers in feature space
• basic assumptions: each cluster pdf has only one maximum; combining pf in mixture function preserrves the maxima
• Algorithm: for each location in feature space find next local maximum with gradien ascent algorithm; if local maximum has no cluster label, assign new label to maximum, add location to cluster

Question 8

SLIC - Simple Linear Iterative Clustering

Answer 8

• supervoxels are group of voxels that share common features
• supervoxels are generated by clustering basedn on intensity similarity and distance
• supervoxel intensity is mean of voxels in corresponding region

Question 9

Kohonens Self-organizing maps

Answer 9

• unsupervised learning with neural networks

- similar to mean shift clustering in attemt to cluster based in inherent attributes