PIBE:A PersonalizableImage Browsing EngineIlaria, Paolo Ciaccia, and M - PDF document

1 PIBE:A PersonalizableImage Browsing Engi
PIBE:A PersonalizableImage Browsing EngineIlaria, Paolo Ciaccia, and Marco PatellaIEIITBO/CNR –DEIS University of Bologna, ItalyCVDB -Paris, June 13, 2004 Bartolini, Ciaccia, and PatellaCVDB –June 13, 20042Why Image Browsing?nExisting Browsing Systems nPIBE (PersonalizableImage Browsing Engine)nCustomizable Hierarchical Browsing StructurenVisual ExamplesnE

2 xperimental EvaluationnConclusions & Fut
xperimental EvaluationnConclusions & Future Work Bartolini, Ciaccia, and PatellaCVDB –June 13, 20043of large image databases (DB’s) is a complex and tedious tasknDifficulties in characterizing the image contentand in defining suitable comparisoninteraction with large DB’s has to include not just queryingbut also browsingto determine a good starting poi

3 nt for searchingnto get an overall view
nt for searchingnto get an overall view of the DB contents Bartolini, Ciaccia, and PatellaCVDB –June 13, 20044Existing Browsing SystemsnSeveral browsing systems existnInformation considered to compute the similarity between images: nautomatically extracted lowlevel features (e.g. color, texture, etc.)nmanually extracted semantic conceptsnTypically, they provide

4 an hierarchical view of the image DBnPro
an hierarchical view of the image DBnProblems that limit their applicability to large image DB’s:based on static browsingstructure nproviding customization facilities that need to reorganizea large part of the DBnno persistentupdating of the browsing structure Bartolini, Ciaccia, and PatellaCVDB –June 13, 20045PIBE (Image Browsing Engine)nA novel adaptiveim

5 age browsing enginenstructure called Bro
age browsing enginenstructure called Browsing Treegraphical personalization actionsto modify the BTn“local” reorganization of the DBnspecific similarity criteria for each portion (subtree) of the BT nuser customizations persistacross different sessions Bartolini, Ciaccia, and PatellaCVDB –June 13, 20046PIBE ArchitecturenPIBE builds off linethe BTfrom

6 the features Browsing Engine Browsing Pr
the features Browsing Engine Browsing Processor At run timethe user interacts with the system by means of a GUIBrowsing Engine:Browsing Processormanages requests to modify the BTndisplays the BT’s content Bartolini, Ciaccia, and PatellaCVDB –June 13, 20047Principles of PIBEnmain ingredientsbehind the BT:nimage descriptors(e.g. color histograms)nas points in

7 a Nfunctions (e.g. weighted Euclidean)n
a Nfunctions (e.g. weighted Euclidean)ninstance of a parametrizedclass of functions to support personalizationclustering algorithms (e.g. kBT is a hierarchical structure derived from a hierarchicalclustering algorithm or, alternatively, by recursivelyapplying a partitioningalgorithm nPIBE is parametricwith respect to the choices of above points. The combination of c

8 hoices should guarantee:nwith respect to
hoices should guarantee:nwith respect to the cardinality of the image DBnsuitable cardinalityof each cluster Bartolini, Ciaccia, and PatellaCVDB –June 13, 20048Browsing TreenPIBE uses:nD HSV color histograms (pweighted Euclidean distances nalgorithm applied to the whole image DB and, recursively, to each of the derived kclusters producing a default BT= 1) neach

9 node of the BT corresponds to a cluster
node of the BT corresponds to a cluster Cof images and maintains thenof Cimage pof Cdefined asnlocal weightvector wcomputed as 22321()øöçèæ=iiipp,jij jjjC);( Bartolini, Ciaccia, and PatellaCVDB –June 13, 20049Browsing Tree Example (k=3) p(C) p(C2) p(C3 ) p(C5) p(C6 ) p(C14) p(C15 ) p(C20) p(C21 p, s q v, uCn = number of images to be clusteredO

10 (kn logk) Bartolini, Ciaccia, and Patel
(kn logk) Bartolini, Ciaccia, and PatellaCVDB –June 13, 200410and Browsing Processorimplement the navigation mechanismnimages displayed in a 2-D screen (spatial visualization technique) using relative distancesnto display images pand qthe weighted Euclidean distance corresponding to the most specific cluster containing both images is usednTwo modalities:n: the

11 user selects a representative image on t
user selects a representative image on the display and zooms inthe cluster contentn: the user explores regions of the space where no representativeimage is present Bartolini, Ciaccia, and PatellaCVDB –June 13, 200411 Browsing Examples Bartolini, Ciaccia, and PatellaCVDB –June 13, 200412 Browsing Examples Bartolini, Ciaccia, and PatellaCVDB –June 13, 20

12 0413Personalization FacilitiesnPIBE prov
0413Personalization FacilitiesnPIBE provides graphicalfacilities allowing the user to move a “source” cluster representative image pon a “target” one poperations nis the “leader”nTwo different actionsn“moving imageis considered as single imagen“moving clusterrepresents the whole cluster C Bartolini, Ciaccia, and PatellaCVDB

13 50;June 13, 200414Updating the BTnBrowsi
50;June 13, 200414Updating the BTnBrowsing Processor is in charge of the BT updates (we consider the more demanding case of moving clusteraction)is recomputed (Ct = Ct È) nis deleted from the BTnalgorithm is recursively applied on the updated Ct = Ct Èusing wt (O(k |representative images of clusters that contained Cneed to be updatedif Cand Chave the same parent C

14 , pdoes not changenif Cand Chave differe
, pdoes not changenif Cand Chave different parents Cand C(but the same grand-parent C), only pand phave to be updated Bartolini, Ciaccia, and PatellaCVDB –June 13, 200415Experimental Evaluationn2,000 color images of IMSI-data setnGoal: “given a set Gof target images (in our experiments, 51 “fishes”), evaluate how PIBE is able to reflect user’

15 s needs in localizing them”Scenario
s needs in localizing them”Scenarios involved:n: this strategy uses the default BTn: a customized version of the BT is considered, following two ways to insert the target images of Ginto a single target cluster Cis an internal node that contains a large number of target images) nis the leaf node containing the highest number of images in G Bartolini, Ciaccia, a

16 nd PatellaCVDB –June 13, 200416Comp
nd PatellaCVDB –June 13, 200416Computed as precision) of the target clusterCwith respect to the set Gof target images nis the the number of target images in Cover the cardinality of C P (%) n. relevant images n. images n. actions down strategy (k=11) 98.08 P (%) n. relevant images n. images n. actions up strategy (k=11) P = | CtG|/| Ct Bartolini, Ciaccia, and Pa

17 tellaCVDB –June 13, 200417Computed
tellaCVDB –June 13, 200417Computed as time(T) of a browsing session, i.e. number of mouse clicks needed to reach all target images by means of the PIBE GUInWe computed the saved(ST) as ST = (1 -where Tis the time needed using a customized BT, whereas Tis that associated to the default BT 63.64 ST (%) T Bartolini, Ciaccia, and PatellaCVDB –June 13, 200418Con

18 clusions & Future WorknPIBE as a new sol
clusions & Future WorknPIBE as a new solution to image browsing:npersonalization graphical actionsnhierarchical structure (Browsing Tree) npersistent and local customizationsnWe are investigating several issues:nvariable fan-out for Browsing Tree nodesn“split” actionsnintegration of semantic concepts extracted by textual descriptors (using lexical ontologie