Parameter Setting and Storage Cost:
We extract SIFT features Star Student Project for each image
in each reference dataset. Based on the local features, we employ
Vocabulary Tree model Star Student Project to build the
initial vocabulary used before the lossy distribution. The
quantization division stops once there are less than Star Student
Project SIFT points in a word node. We denote the hierarchical
level as , Star Student Project which decides the layers to
run the hierarchical clustering. We denote the branching factor as ,
which Star Student Project decides the number of clusters
when a given parent cluster at the higher layer is allowed Star
Student Project to be split into subsets. This setting produces
words in the finest level Star Student Project at most.