Kriengkrai Porkaew, Michael Ortega, and Sharad Mehrotra. Query reformulation for content based multimedia retrieval in MARS. In ICMCS, Vol. 2, pages 747--751, 1999.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....: #RES.a.name,RES.a.address (#BL.o,BL.fv,BL.a (# SI.a.time,SI.a. date (SI) ## NN 5 BL) ## NN 1 RES) 3 Related works During the last decade, several systems that support content based query have been proposed (see the review in [1] Some of the commonly known prototypes are systems such as MARS [4], DISIMA [3] and CHITRA [5] Though many works exist, there are very little of them that consider a multimedia join operation that associates two sets of data for similarity. For example, the MARS system allows complex query formulation by an intelligent query refinement tool for the ....

K. Porkaew, M. Ortega, and S. Mehrotra. Query reformulation for content based multimedia retrieval in MARS. In IEEE International Conference on Multimedia Computing and Systems, volume 2, Florence, Italy, June 1999.

....real number in the [0,1] interval using standard text retrieval techniques. Observe that in our model, in contrast to [39, 36, 25, 38] the data in the database is crisp, i.e. the database does not contain fuzzy terms such as early , small , etc. Traditionally, in a multimedia retrieval model ([23, 27, 11, 18]) the query is represented as a point in multidimensional space, where each dimension is a feature (e.g. a keyword for documents, color and shape for images) Retrieval is performed using a similarity function that measures the distance in multidimensional space between the query point and the ....

....point that represents each object in the database. In contrast, we model the query as a set of similarity functions. Each similarity function is the fuzzy membership function associated with each atom. Further, the multimedia retrieval model considers a weighted sum [27] or the Euclidean distance [18] as the aggregation function to combine query components, whereas we use the fuzzy aggregation function min, as accepted in the fuzzy logic literature (see [33, 11] For approximate retrieval queries it is important to provide the user with iterative and interactive query refinement mechanisms. ....

[Article contains additional citation context not shown here]

Porkaew K., Ortega M., Mehrotra S., Query Reformulation for Content Based Multimedia Retrieval in MARS, in Proceedings of ACM Conference on Multimedia, 1999.

....real number in the [0,1] interval using standard text retrieval techniques. Observe that in our model, in contrast to [33, 30, 21, 32] the data in the database is crisp, i.e. the database does not contain fuzzy terms such as early , small , etc. Traditionally, in a multimedia retrieval model ([19, 23, 9, 14]) the query is represented as a point in multidimensional space, where each dimension is a feature (e.g. a keyword for documents, color and shape for images) Retrieval is performed using a similarity function that measures the distance in multidimensional space between the query point and the ....

....the system automatically refines the query and resubmits it to retrieve a new set of quants. The process continues until the system cannot provide any new quants that satisfy the latest version of the query. Traditionally, methods of query refinement based on user feedback use two techniques ([19, 20]) namely query modification and query re weighting. The first technique involves either query point movement [14] or query expansion [19] The second method automatically adjusts the relative importance of each feature (i.e. weight) to the query. Our approach to query refinement is not using any ....

[Article contains additional citation context not shown here]

Porkaew K., Ortega M., Mehrotra S., "Query Reformulation for Content Based Multimedia Retrieval in MARS", in Proceedings of ACM Conference on Multimedia, 1999.

No context found.

K. Porkaew, M. Ortega, and S. Mehrotra, "Query reformulation for content based multimedia retrieval in MARS," in IEEE Int'l. Conf. on Mult. Comp. Sys., vol. 2, pp. 747-751, 1999. 23

....for the result objects as. A query based on the multiple feature query model can change in two ways: by adding deleting single feature representation queries, and by changing the weights for each single feature query called query reweighting. First we discuss re weighting based on the work in [21] and later how to add delete from the query based on the work in [16] 5.1 Query Re weighting Query re weighting assumes that for the h independent single feature representation queries q i , each with a corresponding weight w i there exist a set of optimal weights w i that capture the users ....

Kriengkrai Porkaew, Sharad Mehrotra, and Michael Ortega, Query Reformulation for Content-Based Multimedia Retrieval in MARS, Proc. IEEE Int. Conf. on Multimedia Computing and Systems (ICMCS), vol. 2, June, 1999.

.... by the user and A is twice as relevant as B, one possible choice of representatives are the points themselves, i.e. h2; f0:4; 0:5; DQ i: There are other possible choices of representatives (see [18] for representative selection techniques) This model was proposed in MARS [18] [19], 20] 2. Reweighting adjusts the dimension weights (i.e. the Q s in (1) to better capture the user s perception within each feature and across features. If the refinement interface is the explicit weight specification interface, the user can modify the weights directly as in Motro s VAGUE ....

....the user. The above refinement step indeed improves the quality of the answers and the answers progressively TABLE 1 Summary of Symbols, Definitions, and Acronyms Fig. 1. Query refinement models. improve with more refinement iterations as demonstrated in references [22] 24] 23] 11] 18] [19], 20] Note that how the query points weights are obtained (i.e. whether we use the VAGUE approach, the MARS approach, or the Mindreader approach) is inconsequential to the discussion in the rest of the paper. This paper discusses how to evaluate the refined query after it has been constructed ....

[Article contains additional citation context not shown here]

K. Porkaew, S. Mehrotra, and M. Ortega, "Query Reformulation for Content Based Multimedia Retrieval in MARS," Proc. IEEE Int'l Conf. Multimedia Computing and Systems, 1999.

....similar images video to the query that is usually expressed by an example image video. The users are also allowed to change the feature weights, paint, sketch, and give feedback. These systems include, but not limited to, QBIC [8] Virage [9] VisualSEEk [10] VideoQ Ill] VIOLONE [12] and MARS [13]. The reader is referred to excellent survey papers, more recently [14] 15] and earlier ones [16] 19] on different aspects of multimedia systems for a more complete review. There is a consensus among researchers on the need to incorporate semantic level representation into these and similar ....

K. Porkaew, M. Ortega, S. Mehrotra, "Query reformulation for content based multimedia retrieval in MAR, S," in IEEE Int'l. Conf. on Mult. Comp. Sys., vol.2, pp. 747-751, 1999.

....feedback of the user on the retrieved objects. Query refinement in MARS consists of query reweighting (QR) and query modification (QM) techniques. QR learns the user s notion of similarity between objects and adjusts the weights of different components of the query. It has been studied in [5, 4]. QM, on the other hand, uses the feedback information to change the query representation to better suit the user s information need. In [5, 2] query point movement (QPM) approach to QM is explored in which a query is represented by a single point in each feature space. At each iteration, the ....

....two ways: query reweighting (QR) and query modification (QM) as described in Section 1. Both refinement mechanisms are combined seamlessly. In this paper, we concentrate only on query modification and the techniques developed can be easily integrated with query reweighting techniques discussed in [4, 5]. MARS supports two query modification approaches as described below. Query Point Movement (QPM) QPM allows only a single object per feature as the query. When the user uses multiple examples to construct the query, the centroid is used as the single point query. Similarly, at each iteration of ....

K. Porkaew, S. Mehrotra, and M. Ortega. Query reformulation for content based multimedia retrieval in MARS. In IEEE ICMCS, 1999.

....F (called representatives [25] in each feature space F 2 F giving rise to multipoint queries (see Figure 1) The weight w F (i) of any representative P (i) F in the multipoint query is proportional to the total weight of all the objects P (i) F represents. This model was proposed in MARS [25, 26, 27]. ffl Reweighting adjusts the inter feature weights (the F s in Equation 3) and the intra feature weights (i.e. the F j s in Equation 1) to better capture the user s perception within each feature and across features. If the refinement interface is the explicit weight specification interface, ....

....the construction of the refined query is complete) the refined query is evaluated and the answers are returned to the user. It has been shown that the above refinement step indeed improves the quality of the answers and the answers gets progressively better with more iterations of refinement [31, 33, 32, 15, 25, 26, 27]. Note that whether the query points weights were explicitly specified or were computed from examples and in the latter case, how exactly the query points (P F and W F ) are obtained (whether MARS or Mindreader approaches) or how the inter feature and intra feature weights are derived are ....

[Article contains additional citation context not shown here]

K. Porkaew, S. Mehrotra, and M. Ortega. Query reformulation for content based multimedia retrieval in MARS. Proceedings of IEEE International Conference on Multimedia Computing and Systems, 1999.

.... , else top is an index node for each child node in top queue.push(N, N) endif enddo Table 3: Single feature query evaluation (The kNN algorithm) the quality of the answers and the answers progressively improve with more refinement iterations as demonstrated in references [22, 24, 23, 11, 18, 19, 20]. Note that how the query points weights are obtained (i.e. whether we use the VAGUE approach, the MARS approach or the Mindreader approach) is inconsequential to the discussion in the rest of the paper. This paper discusses how to evaluate the refined query after it has been constructed using one ....

....of , 4 Note that it is possible to avoid multipoint queries (i.e. support only single point queries) by always using QPM as the query modification technique. However, Porkaew et al. show that QEX based techniques usually perform better than QPM based ones in terms of retrieval effectiveness [19, 18]. Hence supporting multipoint queries efficiently is important for effective and efficient query refinement. Also, since multipoint queries are a generalization of single point queries, supporting multipoint queries makes the techniques developed in this paper applicable irrespective of the query ....

K. Porkaew, S. Mehrotra, and M. Ortega. Query reformulation for content based multimedia retrieval in MARS. Proceedings of IEEE International Conference on Multimedia Computing and Systems, 1999.

.... top is an index node for each child node N in top queue.push(N, MINDIST (QF , N) endif enddo Table 3: Single feature query evaluation (The kNN algorithm) the quality of the answers and the answers progressively improve with more refinement iterations as demonstrated in references [22, 24, 23, 11, 18, 19, 20]. Note that how the query points weights are obtained (i.e. whether we use the VAGUE approach, the MARS approach or the Mindreader approach) is inconsequential to the discussion in the rest of the paper. This paper discusses how to evaluate the refined query after it has been constructed using one ....

....of RN , 4 Note that it is possible to avoid multipoint queries (i.e. support only single point queries) by always using QPM as the query modification technique. However, Porkaew et al. show that QEX based techniques usually perform better than QPM based ones in terms of retrieval effectiveness [19, 18]. Hence supporting multipoint queries efficiently is important for effective and efficient query refinement. Also, since multipoint queries are a generalization of single point queries, supporting multipoint queries makes the techniques developed in this paper applicable irrespective of the query ....

K. Porkaew, S. Mehrotra, and M. Ortega. Query reformulation for content based multimedia retrieval in MARS. Proceedings of IEEE International Conference on Multimedia Computing and Systems, 1999.

....Analysis and Retrieval System (MARS) project to provide retrieval capabilities to rich multimedia data. Our research addresses several aspects of multimedia information retrieval including extraction of features from multimedia object that model its content [16] multimedia object and query models [20]; multimedia retrieval models that define how similarity between a multimedia object and the query can be computed [17, 20] query refinement techniques that modify the query representation based on the relevance feedback provided by the user to better meet the users information need [20, 19, 5] ....

.... several aspects of multimedia information retrieval including extraction of features from multimedia object that model its content [16] multimedia object and query models [20] multimedia retrieval models that define how similarity between a multimedia object and the query can be computed [17, 20]; query refinement techniques that modify the query representation based on the relevance feedback provided by the user to better meet the users information need [20, 19, 5] efficient algorithms to evaluate content based retrieval of multimedia objects; efficient techniques to implement query ....

[Article contains additional citation context not shown here]

K. Porkaew, S. Mehrotra, and M. Ortega. Query reformulation for content based multimedia retrieval in MARS. In Proc. IEEE Conf. on Multimedia Computing and Systems, Florence, Italy, June 1999.

No context found.

K. Porkaew, S. Mehrotra, and M. Ortega, #Query reformulation for content based multimedia retrieval in MARS," in Proc. IEEE Conf. on Multimedia Computing and Systems, #Florence, Italy#, June 1999.

No context found.

K. Porkaew, S. Mehrotra, and M. Ortega, "Query reformulation for content based multimedia retrieval in MARS," in Proc. IEEE Conf. on Multimedia Computing and Systems, (Florence, Italy), June 1999.

....feature as shown in Figure 2. In this paper, we focus on the different approaches of query refinement at the individual feature level and query evaluation techniques for these approaches. These approaches can be easily integrated into query refinement involving multiple features as proposed in [15]. At the individual feature level, MARS supports two query modification 3 (a) Query Point Movement (b) Query Expansion Figure 3. Query Refinement Approaches approaches: 1) Query Point Movement and (2) Query Expansion. We discuss them in detail below. 3.1 Query Point Movement Query point ....

.... of interactive image database annotation in Minka and Picard [23] interactive mechanisms for searching images on the world wide web in ImageRover [19] and interactive integration of keywords and high level concepts to improve the retrieval effectiveness in WebSEEk [20] In our previous work [17, 18, 15], we introduced relevance feedback as a technique to capture the subjective element in multimedia retrieval systems. Our experiments has shown that this is a promising research direction. We use the feedback information to modify the query and adjust the weights. For example, within a feature, we ....

K. Porkaew, S. Mehrotra, and M. Ortega. Query reformulation for content based multimedia retrieval in MARS. In IEEE Int'l Conf. on Multimedia Computing and Systems, Florence, Italy, June 1999.

....mechanisms that can efficiently support query expansion using multidimensional index structures. 1 Introduction In a content based multimedia retrieval system, it is difficult for users to specify their information need in a query over the feature sets used to represent the multimedia objects [10, 7, 12]. Motivated by this, recently, many content based multimedia retrieval systems have explored a query by example (QBE) framework for formulating similarity queries over multimedia objects (e.g. QBIC [4] VIRAGE [1] Photobook [9] MARS [6] In QBE, a user formulates a query by providing examples ....

....of the different features used to represent the multimedia objects to the query. To overcome the above limitations, in the Multimedia Analysis and Retrieval (MARS) project, we explored techniques that allow users to refine the initial query during the retrieval process using relevance feedback [10]. Given an initial query, the system retrieves objects that are most similar to the query. The feedback from the user about the relevance of the retrieved objects is then used to adjust the query representation. Relevance feedback in MARS serves two purposes as follows. Query Reweighting adjusts ....

[Article contains additional citation context not shown here]

Kriengkrai Porkaew, Sharad Mehrotra, and Michael Ortega. Query reformulation for content based multimedia retrieval in MARS. In IEEE Int'l Conf. on Multimedia Computing and Systems, 1999.

....centroid. Effectively, for each F , the query point is moved towards the relevant objects [30, 32, 31, 13] ffl Query Expansion (QEX) In this model, the relevant objects are represented by multiple points (called representatives [24] in each feature space giving rise to multipoint queries [24, 25, 26]. The weight (wF i s) of any representative R in the multipoint query is proportional to the total weight of all objects R represents. ffl Intra Feature Reweighting: which dynamically determines, for each feature space F 2 F , the distance function DF that best suits the user s perception for ....

....perception for that feature by adjusting the intra feature weights (i.e. the F j s in Equation 1) based on her feedback. The three techniques 4 are combined seamlessly in the refinement model. As the query gets progressively refined, the quality of answers returned gets progressively better [30, 32, 31, 13, 24, 25, 26]. The specifics of how reweighting is achieved at inter feature and intra feature levels and how the query point is changed in QPM or how new points are added in QEX are inconsequential to the discussion in the rest of the paper. The techniques developed here will work irrespective of the exact ....

[Article contains additional citation context not shown here]

K. Porkaew, S. Mehrotra, and M. Ortega. Query reformulation for content based multimedia retrieval in MARS. Proceedings of IEEE International Conference on Multimedia Computing and Systems, 1999.

No context found.

Kriengkrai Porkaew, Michael Ortega, and Sharad Mehrotra. Query reformulation for content based multimedia retrieval in MARS. In ICMCS, Vol. 2, pages 747--751, 1999.

No context found.

K. Porkaew, S. Mehrota, and M. Ortega. Query reformulation for content based multimedia retrieval in mars. ICMCS, pages 747-751, 1999.

No context found.

K Porkaew, M Ortega, and S Mehrotra. Query reformulation for content based multimedia retrieval in MARS. In ICMCS, Vol. 2, pages 747--751, 1999.

No context found.

K. Porkaew, S. Mehrota, and M. Ortega. Query reformulation for content based multimedia retrieval in mars. ICMCS, pages 747-751, 1999.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC