Furthermore, we should have a histogram where the orienatations (0-180 degrees) of each line are reflected. We will break the orientations into 8 bins, each 22.5 degrees, which then shows a probability density function. So, instead of saying that a bin containx "n" lines, we would give the computed value n/number of lines in the cluster.
For the image: image size number of clusters For each cluster: number of lines area density = number of lines / area relative area = area / image size
The idea behind all of this is that we want to study the behavior of this classifier and figure out the trade-off between generalization and error.