Self-adaptation in Image and Video Retrieval - Multimedia Database Retrieval: Technology and Applications - page 60

Traditionally, the evaluation of pseudo labels is carried out on the feature space
in a (single) domain similar to the feature space used for image ranking. Given
the feature space F 1 , an initial ranking set is obtained for retrieval. Then, the

system assumes that a small number of top-ranked objects in the initial set are
the pseudo positive samples, and the latest-ranked objects are given the pseudo
negative samples [ 52 , 53 ]. This rule for labeling data uses the query sample (labeled
data) to explore unlabeled data in order to increase the size of training set. This can
be viewed as the transductive learning problem, which has been studied to handle
small numbers of labeled data [ 54 - 57 ]. The pseudo-labels are inferred using the
nearest-neighbor rule applied to the unlabeled samples in the dataset. This method
enlarges training sample sets, and has been used to improve effectiveness of the
support vector machine (SVM)-based adaptive retrieval [ 52 , 58 , 59 , 307 ].
There is a difficulty in making assumptions about the class assigned to unlabeled
data. The nearest-neighbor rule applied on the feature space

F 1 , which is the same
as the feature space for obtaining initial ranking, can lead to imprecision in class
information. The top-ranked samples are not always the relevant, correct answers
that meet the user's information needs, due to the limited accuracy of current
multimedia retrieval systems [ 60 ]. Alternatively, instead of using ranking scores
on a single domain for the assignment of pseudo labeling, the self-organization
methods [ 61 ] can be adopted for this task on a different feature space F 2 .The

systems label the unlabeled points according to the clusters to which they naturally
belong. An advantage of the self-organizing method is that it may be able to make
better predictions, with fewer labeled points than standard pseudo-RF, because it
uses the natural breaks found in the unlabeled points.
Figure 3.1 illustrates the process for assignment of pseudo labeling using
two feature sets in F 2 , which is used for the adaptation of image/video

retrieval system. The feature set in

F 1 and

F 1 is a standard feature used for ranking
image/video database, whereas the feature set in F 2 has high quality features

to be used for relevant judgment (the assignment of pseudo labels) by the self-
organization methods. Both sets can be characterized by visual descriptor, text,
and other modalities of multimedia files in the database. For instance, previous
works [ 56 , 62 - 65 ] have used pseudo-positive samples in the visual domain for query
expansion in the text domain. This is possible due to the availability of metadata
associated with images, especially for web image retrieval applications [ 66 , 67 ].
In order to perform the assignment of pseudo labeling on the feature set

F 2 ,
the self-organization method SOTM is adopted. The motivation is that this method
is suitable for clustering sparsely distributed data in the current application. The
feature space F 2 is usually of high dimension and only a small number of training

samples is considered for the assignment of pseudo labels. The efficiency and
flexibility of the SOTM in adapting to, and its implicit awareness of topology of
input space, make it an appropriate candidate for implementing this idea.
The SOTM [ 68 ] attempts to partition a feature space description of input data,
by locating clusters of high density within this feature space. Competitive learning
is used to locate clusters such that the final representation maintains the general
topology of the feature space, yet doing so in a flexible and efficient manner by