Instead of using clustering methods to group images according to the distances measured we started with an approximation based on a threshold distance, since applying a threshold to the ordered series of distance values is simple to implement and gives an idea of what better clustering algorithms would produce. Our approach can be considered the inverse of that followed by [ 3 ] to get an overview of videos by generating key-frames based on a threshold setting. Starting with the commonest first frame, also first index frame, all subsequent images within the threshold distance are written to the cluster loop for the first index frame while eliminating all row- and column entries associated with the cluster frames; the first image further away then the threshold distance becomes then next index frame and all images within the threshold distance are written to its cluster loop; this is repeated until there are no images left.
The overview of the image database in figure 1 starts up with the index frame loop, which can be played off at different speeds or frame by frame in both directions; when the loop display is stopped, the cluster loop is displayed by clicking an index frame. This way one can move quickly between very different classes of images or concentrate ones attention to within a certain class of images by displaying index frame associated cluster loops.
Dr. D.P. Huijsmans