Mobile Landmark Recognition - Multimedia Database Retrieval: Technology and Applications - page 136

where N denotes the total number of images in the database, N v i denotes the number

of images containing v i , and log N v i is the inverted term frequency of v i in the
entire dataset.

5.5 Re-ranking Approach to Landmark Recognition

This section presents a re-ranking method for improving the performance of the
recognition system employing SVT and BoW. The re-ranking method selects a
suitable weighting scheme by applying an unsupervised wrapper feature selection
to a training set associated with a particular query. The method is described in the
following sections.

5.5.1 Building a Training Set via Ranking

Given a known database containing N landmark photos [ I 1 , I 2 ,..., I N ] , the first step

in the learning process is to obtain a subset of photos [ A 1 , A 2 ,..., A R ] to build a

training set for a given query I q .Here R N , and all landmark images in the

database are available with the class labels (or the information on the landmarks).
Based on the distance function D d ( I x , I q ) , this step outputs the following ranking

list:

Query ( I q )=[ A 1 , A 2 ,..., A R ] (5.20)

where A i is the i -th returning entry of the query. [ A 1 , A 2 ,..., A R ] are R top ranked

images based on the original BoW histogram given the query.
We aim to maximally improve the ranking order by modifying the BoW of the
query instead of using that of the original query.

5.5.2 Unsupervised Wrapper Feature Selection Method

As a specific subset of features is specially effective for the accurate prediction of
certain query classes, the original BoW features can be modified by this subset. The
unsupervised wrapper method [ 163 ] is utilized for feature selection. This process is
shown in Fig. 5.2 . The ranked BoW vectors can be first divided into subgroups by
the unsupervised clustering. The single linkage (SL) [ 164 ] is selected to accomplish
this purpose. Once each of the subgroups are artificially assigned a class label, then
the wrapper method is applied.