Segmentation Flashcards
Segmentation and grouping
Components on representation have similar visual properties
Tokens
Whatever we need to group
Goals of image segmentation
Meaningful segmentation of an image
Pixels to regions changes the representation
Algorithm can be tested quickly
Cannot be expected to work well in some cases
Classic methods of segmentation
Thresholding
Split and merge
Segmentation by clustering
Tokens are visually similar to one another and are grouped as such
Types of clustering
Hierarchal clustering
K-means clustering
mean-shift clustering
Spectral clustering
Hierarchal clustering
Agglomerative- bottoms up, start with one token and then add to cluster
Divisive- top-down(splitting)
Start with all tokens in one cluster than split or take out tokens from cluster
Issues with hierarchal clustering
Determine a good inter cluster distance
How many clusters are there?
K-means clustering
Give a set of vectors
Specify k number of desires clusters
Divides into clusters to minimize sum of distance between elements in clusters
K-means issues
Initial cluster estimate could be bad, how do we determine a good number
Clusters found by k-means tend to be spherical
Mean shift clustering
Clusters are places where data points tend to be close together
Searched for modes and adds points to clusters
Mean shift pros and cons
Pros- model free Single parameter (window size) Robust to outliers
Cons Output depends on window size Window size selection is not trivial Computationally expensive Does not scale
Images as graphs
Node for every pixel
Edge between every close pixel
Each edge weighted by similarity
Segmentation by graph
Break graph into segments based on
Links that cross between segments
Have low affinity/similarity
Does scale affect affinity?
Yes
Spectral clustering algorithms
Pre-processing
Construct a matrix
Decomposition
Compute eigenvalues
Grouping
Assign points to two or more clusters
Spectral advantage
Spectral space representation is easy to understand
Spectral clustering problem
Eigenvalues of an affinity matrix can be misleading to guide clusters
Drawbacks of cuts based on minimum weight
Favors cutting small sets of isolated nodes
Normalized cut algorithm
Compute matrixes W D
Solve for eigenvectors
Use eigenvector with smallest Eugene value to partition the graph
Decide if current partition needs to be subdivided further
Normalized cuts pros and cons
Pros- generic framework and can be used with many different formulas
Cons- high storage requirement
Bias towards partitioning
