Home People Research Publications Demos
News Jobs Prospective
About Internal

Examples for query using QBSE and QBVE

Some examples of Semantic Multinomial

Quering using single image on Corel50

Some examples where QBSE performs better than QBVE. The second row of every query shows the images retrieved by QBSE. Note, for example, that for the query containing white smoke and a large area of dark train, QBVE tends to retrieve images with whitish components, mixed with dark components, that have little connection to the train theme.

Quering using multiple image on Flickr18

Examples of multiple-image QBSE queries. Two queries (for Township and Helicopter) are shown, each combining two examples. In each case, two top rows presents the single-image QBSE results, while the third presents the combined query. It illustrates the wide variability of visual appearance of the images in the Township class. While single-image queries fail to express the semantic richness of the class, the combination of the two images allows the QBSE system to expand indoor market scene and buildings in open air to an open market street or even a railway platform. This is revealed, by the SMN of the combined query, presented below, which is a semantically richer description of the visual concept Township, containing concepts (like sky, people, street, skyline) from both individual query SMNs.

Detailed Example showing why QBSE works.

Query from class commercial construction with top QBSE and QBVE matches shown. For QBSE, below each image are also published the semantic features of highest posterior probability. The semantic features of largest probability include various words that are clearly related to the concept of construction. Outside the semantic space, retrieval success is purely due to the effectiveness of contextual relationships.

Back to QBSE