(AKM) [5], Robust Approximated K-Means (RAKM) [18]
and Approximate Gaussian Mixture (AGM) [19] lead to
heavy training and retrieval time overhead. Additionally, in
the index method, which is based on a vocabulary tree
[20–22], the scale (the number of branches and levels) of
a vocabulary tree is needed to ensure the image matching
precision. While the image training set is huge (say
millions or billions), it is almost impossible to train such
a huge vocabulary tree in centralized memory. Moreover,
the centralized training time of a huge vocabulary tree is
too long.
In this paper, in order to reduce the memory and time
overhead for centralized training of a huge vocabulary tree so
that it can train a huge image training set, a new distributed
vocabulary tree image training and retrieval scheme is proposed.
The major contributions of this paper are: