clustering

group together similar data points, usually in feature space

Clustering pixels

Top-down:

• pixels belong together because they lie on the same visual entity (object, scene…)

Bottom-up:

• pixels belong together because they look similar
• represent pixels in a feature space (e.g. RGB + XY)
• define pair-wise distance metric
• group pixels by similarity Distance metrics
• euclidian - straight-line in feature space
• cosine sim - angle between vectors (good for high-dim data)

algos

• k-means
• heirarchical agglomerative
• mean shift
• normalized cuts (graph-based)

how to choose feature space? can group pixels in diff ways

• color sim (3-dim)
• edges and blobs
  • oriented gradient sim
• RGB + XY location
  • intensity+position sim

results don’t have to be spatially coherent

• clustering on rgbxy makes it more spatial coherence

---