Home People Research Publications Demos
News Jobs Prospective
About Internal

Query by Semantic Example (Demo)

Two main content based image retrieval paradigms have evolved over the years: one based on visual queries, or query-by-visual-example (QBVE), and the other based on textual queries, or semantic retrieval . Early retrieval architectures were almost exclusively based on the visual retrieval framework. Under this paradigm, each image is decomposed into a number of low-level visual features (e.g. a color histogram), and retrieval implemented through query-by-example. This consists of specifying the query as the feature vector extracted from an image, and searching the database for the best match to that feature vector. It was, however, quickly realized that strict visual similarity is, in most cases, weakly correlated with the similarity criteria adopted by humans for image comparison.

This motivated the more ambitious goal of designing retrieval systems with support for semantic queries. The basic idea is to annotate images with semantic keywords, enabling users to specify their queries through natural language descriptions. This idea was extended to the design of semantic spaces, i.e. spaces whose dimensions are the semantic concepts known to the retrieval system. The earliest among such systems were based on semantic information extracted from image metadata. Later on, semantic spaces were also constructed with resort to active learning, based on user relevance feedback. Because manual supervision is labor intensive, much of the research turned to the problem of automatically extracting semantic descriptors from images, by application of machine learning algorithms. Early efforts targeted the extraction of specific semantics. More recently, there has been an effort to solve the problem in greater generality, through the design of techniques capable of learning relatively large semantic vocabularies from informally annotated training image collections.

When compared to QBVE, semantic retrieval systems have the advantages of 1) a higher level of abstraction, and 2) easier query specification (through the use of natural language). They are, nevertheless, restricted by the size of their vocabularies, and the fact that most images have multiple semantic interpretations. None of these limitations afflict QBVE, which also tends to enable more intuitive implementation of interactive functions, such as relevance feedback. In this work we show that, through the simple extension of the query-by-example paradigm to the semantic domain, it is possible to design systems that combine the benefits of the two classical paradigms. The basic idea is to 1) define a semantic space , where each image is represented by the vector of posterior concept probabilities assigned to it by a semantic retrieval system, and 2) perform query-by-example in this space. We refer to this combination as query-by-semantic-example (QBSE), and present an extensive comparison of its performance with that of QBVE.

Semantic image retrieval. Left: Under QBSE the user provides a query image, probabilities are computed for all concepts, and the image represented by the concept probability distribution. Right: Under the traditional SR paradigm, the user specifies a short natural language description, and only a small number of concepts are assigned a non-zero posterior probability.

Demo On Corel dataset using 104 annotations
Results: Quantitative results (Perfomance measures)
Some examples of image queries
Presentation: Query by semantic example.ppt
Databases: We have used the following data-sets for image retrieval experiments. Please contact the respective people for information about obtaining the data: The images from these data-sets (except Flickr18, which are from the website Flickr.com ) are from the Corel image CDs. The annotations for Corel50 are a subset of those from the Berkeley Digital Library project. The full set of annotations for 40,000 Corel images is available here.

Publications: Bridging the Gap: Query by Semantic Example
Rasiwasia, N., Moreno, P. L., Vasconcelos, N.
Multimedia, IEEE Transactions on,
Vol. 9(5), pp. 923-938, Aug 2007.
  Query By Semantic Example
Nikhil Rasiwasia, Nuno Vasconcelos, Pedro J Moreno
Proceedings of the International Conference on Image and Video Retrieval LNCS 4071, pp. 51-60
Phoenix, Arizona, 2006.
Contact: Nuno Vasconcelos, Nikhil Rasiwasia