advanced clustering-based segmentation technique where mean shift moves towards dense regions in the data

  1. represent each pixel w feature vector
  2. define window over pixel
  3. identify all pixels within radius of feature vector
  4. compute mean of all pts inside window (center of gravity among neighbors of window)
  5. shift window towards mean feature location
  6. repeat until convergence (reaching high-density region)

  • initialize multiple windows at random
    • pixels that end up in same location belong to same cluster
  • attraction basin: feature region for which all windows end up in same location
    • speeds up

pro

  • no need to specify K
  • better at irregular cluster shapes than k-means
  • does not assume prior shape
  • single parameter
  • robust to outliers

cons

  • expensive - need to shift one window for every pixel - redundancy
  • window size is hyperparameter that is hard to find
  • does not scale well with dim of feature space