Clusterize the Clusters
In the above matrix illustration, one can see the CK1 distances used to compute the hierarchical clustering graph displayed below. Indeed, one has 8 different images, labeled 1 to 8 in the below graph.
We now use 5 images that we consider to be very similar to build a new tree and see what distances we get, as well as what hierarchy we obtain using the same process.
Using the information from this tre, we find that our threshold T is probably higher than 0.6.
In this tree, image 7 is grouped with images 1 to 5 before being grouped with image 6, which is undesirable. We think this is due to the fact that the sizes of images 1-5 and 6-7 are too different and the resizing has a negative effect.