• group pixels w a segmentation algorithm
    • can group pixels by optical flow - “pixels that move together are part of the same entity”

Unsupervised segmentation: how can pixels in an image be grouped w/o prior labels?

Use cases

  • object identification
  • efficiency (processing groups rather than individual pixels)
  • foreground-background separation

Examples

Challenges:

  • noise
  • blur
  • color variation

What is a good segment?

  • proximity (nearby pixels belong together)
  • similarity (color, shape)
  • common fate (moving together)
  • symmetry & continuity (aligned edges )

But it is subjective and context-dependent.

  • gesalt theory: whole is greater than sum of its parts - relationships btw parts can yield new properties
    • grouping is key to visual perception

next: clustering