Interactive Search by Direct Manipulation of Dissimilarity Space

Abstract

Abstract-In this paper, we argue to learn dissimilarity for interactive search in content based image retrieval. In literature, dissimilarity is often learned via the feature space by feature selection, feature weighting or by adjusting the parameters of a function of the features. Other than existing techniques, we use feedback to adjust the dissimilarity space independent of feature space. This has the great advantage that it manipulates dissimilarity directly. To create a dissimilarity space, we use the method proposed by Pekalska and Duin, selecting a set of images called prototypes and computing distances to those prototypes for all images in the collection. After the user gives feedback, we apply active learning with a one-class support vector machine to decide the movement of images such that relevant images stay close together while irrelevant ones are pushed away (the work of Guo et al.). The dissimilarity space is then adjusted accordingly. Results on a Corel dataset of 10000 images and a TrecVid collection of 43907 keyframes show that our proposed approach is not only intuitive, it also significantly improves the retrieval performance.Index Terms-Active learning, dissimilarity learning, interactive image search, visualization.

Towards interactive, intelligent, and integrated multimedia analytics

“…Intermediate Advanced I-SI [85] Newdle [93] sVisit [56] Informedia [33] Canopy [8] Similiarity browser [61] INA browser [81] MediaTable [19] [20] semantic gap pragmatic gap Fig. 8.…”

“…Being continuously updated since then, Informedia has received a relevance feedback component in 2008 [33] (albeit one refining search results, rather than maintaining an adaptive model). Another early pioneer system is the similarity manipulation browser by Nguyen et al [61]. This approach employs a similarity space browser through which the user directly manipulates the similarity space, with the machine recomputing the used similarity and rearranging the items based on the interactions.…”

Grouping real functions defined on 3D surfaces

“…For instance the AdaBoost algorithm [25] has been used in [20] to select relevant views of 3D objects with respect to the light field descriptor and in [26] to automatically select the frequencies of the Laplacian spectrum of 3D shapes that are relevant for shape classification. Also, the tuning of the descriptors used to the specific context or user needs has been addressed implicitly by relevance feedback techniques [27,28].…”

A Survey on Visual Content-Based Video Indexing and Retrieval

“…Then, the videos closest to the classifier boundary are returned to users for identification and the system is updated using the identified videos. Nguyen et al [270] also use active learning in the interaction process to choose videos close to the classifier boundary. In contrast with the aforementioned algorithm that selects videos in the feature space, they choose videos in the dissimilarity space represented by a number of prototypes.…”