CP467 - Clustering

What is clustering?

Grouping together similar points representing them with a single token

What is the goal of K-Means clustering?

Choose three "centres" as the representative intensities and label every pixel according to which of these centres it is nearest to

What is the K-Means clustering algorithm?

1. Randomly initialize k points to act as cluster centres
2. Given cluster centres, determine points in each cluster
3. Given points in each cluster, solve for cj
4. If cj have changed, repeat from Step 2

What can we base grouping pixels on?

1. Intensity
2. Colour
3. Texture

What are some pros of K-means clustering?

1. Simple and fast
2. Easy to implement
3. Converges to local min of within-cluster squared error

What are some cons of K-means clustering?

1. Specifying k
2. Sensitive to initial centres
3. Sensitive to outliers
4. Detects spherical clusters only

What are the steps in the mean shift clustering algorithm?

1. Centre a window on that point
2. Identify all points in the window then compute their mean
3. Centre the window at the new mean location
4. Repeat (2,3) until convergence

What are some pros of mean shift?

1. Generic technique
2. Does not assume any prior shape on data
3. Can handle arbitrary feature spaces
4. Only one parameter to choose
5. h has a physical meaning, unlike K-means
6. Robust to outliers

What are some cons of mean shift?

1. Window size is not trivial
2. Does not scale well with dimension of feature space