Segmentation and Clustering p3 - Lecture 9 - Week 4

Question 1

What is mean-shift segmentation?

Answer 1

An advanced and versatile technique for clustering-based segmentation

Question 2

What is a mode in a histogram?

Answer 2

Mode = local maximum of the density of a given distribution

Question 3

What is iterative mode search?

Answer 3

1. Initialise random seed, and window W
2. Calculate centre of gravity (the “mean”) of W
   
   • Sum of x in W (x * Hist(x))
3. Shift the search window to the “mean”
4. Repeat step 2 until convergence

Question 4

What is a cluster in Mean-Shift clustering?

Answer 4

All data points in the attraction basin of a mode

Question 5

What is an attraction basin in mean-shift clustering?

Answer 5

The region for which all trajectories leads to the same mode

Question 6

How does mean-shift achieve clustering/segmentation?

Answer 6

Find features (colour, gradients, texture, etc…)
Initialise windows at individual pixel locations
Perform mean shift for each window until convergence
Merge windows that end up nea the same “peak” or mode

Question 7

Mean-Shift pros and cons

Answer 7

Pros
- General, application-independent tool
- Model-free, does not assume any prior shape (spherical, elliptical, etc…) on data clusters
- Just a single parameter (window size h)
- Finds variable number of modes
- Robust to outliers

Cons
- Output depends on window size
- Window size (bandwidth) selection is not trivial
- Computationally (relatively) expensive (~2s per image)
- Does not scale well with dimension of feature space