Segmentation of Parallax Maps Using Clustering

Problem: segment parallax map according to surface orientation.

3-D Plots of parallax maps (left cone, right cylinder):

Procedure:

  1. Take parallax map.
  2. Decrease measurement noise: Filter map using 3x3 average filtering.
  3. Determine local surface orientation: Take 3x3 pixel window from image -> subtract neighboring pixels from center pixel -> form 8-dimensional vector.
  4. Determine local surface orientation for each pixel of map.
  5. Train 1-dimensional Kohonen Self-Organizing Feature Map.
  6. Each weight vector of processing element correspond to certain way oriented surface.
  7. Search closest weight vector for each 8-D vector -> Assign vector to corresponding cluster.
  8. Form image about clustered vectors.
Results of segmented cone: Top left: parallax map; top right: segmentation result, 5 groups; bottom left: segmentation result, 7 groups; bottom right: segmentation result, 9 groups.

Results of segmented cylinder: Top left: parallax map; top right: segmentation result, 9 groups; bottom left: segmentation result, 11 groups; bottom right: segmentation result, 15 groups. Some of the groups are so small that they are not visible in segmented image.