Database Reference
In-Depth Information
Fig. 5.5
Samples images used for (
a
)trainingand(
b
) testing in the Art Science Museum category
Table
5.2
shows the comparison of the percentage accuracy for the World
landmark database. The highest score was 77.25 % at the 90:10 ratio, obtained
by the SA + SVT + Re-ranking method. The results observed for the Singapore
landmark database can also be observed for the World landmark database with
regard to the performance of the methods being compared. The system which
implemented the re-ranking method outperformed the baseline recognition system
which implemented the SA and SVT methods.
5.7
Summary
The chapter extends the conventional Bag-of-Words (BoW) method for image
content indexing to a discriminating BoW method for landmark recognition. Both
scalable vocabulary tree (SVT) and BoW representation can be applied with
the nonlinear discrimination power, by taking into account saliency weighting
in image recognition. The discriminating BoW model is obtained by the feature
selection and re-ranking processes, based on an unsupervised wrapper approach.
This feature selection method has a great effect on learning efficiency and/or
prediction performance.
Search WWH ::
Custom Search