Vision

Overview

• Input to (monochrome) vision system: 2-dimensional array of brightnesses (intensities)
• Output of vision system: list of recognized objects and their relative locations, physical properties, and relations to each other
• Edge detection
• Depth
• Shape from shading
• Line labeling
• Object recognition, relation recognition

Edge Detection

• The importance of edges: boundaries between objects; borders of objects; changes in reflectance, illumination, depth, and surface orientation
• The problem of noise: smoothing
• Smoothing by convolving with a point-spread function
• Finding an edge by taking the slope of the slope (the second derivative) of brightness and looking for zero crossings
• Combining smoothing with edge detection by convolving with a point-spread function that averages and finds the slope of the slope
• Labeling edge fragments for orientation using cooperative (relaxation) algorithms
  • Set of elements to be interpreted and a set of constraints between adjacent elements
  • Algorithm repeatedly adjusts the interpretation of each element to be in greater harmony with neighboring elements (constraint satisfaction/propagation)
  • Exploit parallelism: one processor per element; processors need only talk to their neighbors

Depth

• Stereo disparity: requires determining correspondences between points in images
• Movement of objects
• Movement of the eyes
• Shading
• Knowledge about occlusion
• Knowledge about relative size

Line Labeling

• Real edges -- Interior: convex, concave (various types), boundary: 2 ``directions''
• Pseudo-edges -- marks, shadow edges, cracks
• Juncftion types
• Arriving at possible labeled junctions by viewing figures composed of one, three, five, and seven cubes from different angles
• Possible labeled junctions: 18 (out of 208)
• Labeling beginning with borders, using constraints on arrow and fork junctions