Image segmentation methods often operate with the goal of partitioning the image into regions of uniform colour. Because real-world objects can have many different colours, we hypothesize that a segment uniformity measure based on colour distributions, instead of mean colour for example, is more appropriate for object segmentation. To demonstrate this, a histogram comparison metric known as the Earth Mover’s Distance is used with a graph contraction algorithm in a new segmentation technique. Graph contraction is used for its efficiency and its ability to generate multiple-level segmentations called pyramids. The advantages of pyramidal object segmentation for object recognition and stereoscopy, two prevalent problems in computer vision, are discussed. Output is shown for images from the Berkeley Segmentation Data Set and compared against manual image segmentations.
Full Text: PDF