Home

User`s Manual Release 1.2

1. 30 8 3 1 Gaussian mixture 31 8 3 2 Kmeans 4 32 8 3 3 Fuzzy kmeans 33 8 3 4 Density based classification DBSCAN 33 8 3 9 Minimal spanning tree and hierarchical classification 34 8 4 Spatial reconfiguration 39 8 4 1 Comparison 39 8 4 2 Amma aaa a 36 8 4 3 Move 2 a 37 8 4 4 Observation axis 37 8 5 Explore from Images 2 a 37 Explorer3D User s manual 3 86 SVM 2 4444 ee 38 8 7 kemels 44444444 41 8 8 2D Graphs a 42 8 9 Network driven control 4 4 4 42 8 10 Data import 4 44 9 Secondary tools 46 9 1 Travel throughout dimensions 46 4 M Exbrayat L Martin 1 A Foreword on the english documentation Explorer3D has been first written in French So was its documentation The software has been for long internationalized The documentation has been more recently Most of the snapshots come from the French version They might thus slightly differ from the actual english version interface If you feel like updating the snapshots or improve the english level of this documentation do not hesitate to contact us 2 What is Explorer3D Explorer3D is an interactive visualization software It allow
2. Fichiers Version de base 1 2 3 4 IE 1 name __ MDVP Fo MDVP Fhi MDVP Flo Na 2 jphon RO 119 99200 157 30200 74 99700 G tape 3 Type de donn es 3 phon RO 122 40000 148 65000 113 81900 one Facts a 4 phon_RO 1116 68200 131 11100 111 55500 d anes A 5 a phon Ro 116 67600 137 87100 111 3600 d A ons dobeta 6 phon RO 116 01400 141 78100 110 65500 g Une matrice de distances Z _ phon Ro 120 55200 131 16200 113 78700 8 phon RO 120 26700 137 24400 114 82000 g phon RO 107 33200 113 84000 104 31500 1 phon_ RO 95 73000 132 06800 91 75400 q ae woo ne oe eoo a EE ji Nouvelle version Figure 40 Import data type 8 outil d importation de donn es Fichi Iculer la projection sent une information additionnelle ensemble des lignes de 2 Ja fise Noms d attributs Optionnel jer sour Si votre fichi ce contient le nom des descripteurs indiquer le num ro de la ligne 1 valdes Figure 41 Import description of file content a view of the final imported file using the Explorer3D format fig 43 The user can then either save the file or save it and compute a projection Figure 44 presents the PCA computed after importing the Parkinson dataset 8 Outil d importation de donn es Fichiers Version de base E 1 a name o o 3 J 4 al P Fhi M
3. To be translated Several kinds of input file can contain data for only a subset of the objects This does mainly make sens in multisource mode were some input can be avail able for a subset only of the objects but were the user wishes to compare the resulating projection with a former one based on another input data file A subset file is denoted by the presence of the SUBSET keyword see the various file formats The global rank of each object must then be given In the current release the user 1 supposed to load at least one complete set before loading a subset Subset files can contain additional data columns The value of these columns will be set to UNDEFINED for the missing objects 5 4 Features input file Input file do commonly consist of a set of features i e a table where each object is described according to a set of features A 3D view is then computed by projecting the objects in a low dimension space that reflects the original large dimension space formed by the original features 5 4 1 General format Explorer3D accepts text files structured as follows Explorer3D User s manual 7 SUBSET START WITH x number of objects number of original features excluding additional features names of original features followed by names of additional features Description of objects list of original and additional features one object per line Feature names and values are separated by a delimiter which is b
4. Afficher les images hors 3D Calcul de la vue 3D au chargement pa Zoomer dans une nouvelle fen tre Activation du systeme d envoi distant Le Affichage des objets sources tableau L Affichage dynamique des images Y Vision st r oscopique si disponible m Taille des labels N LE Taille des objets dynamique lv Taille des objets par d faut Commande via le r seau N de port L gende activ e lv OK Figure 16 Options 7 8 Legend Depending on the object colouring policy the legend window displayed will vary The standard legend window one colour per object a finite set of colours as already been presented in section 6 1 Two others kinds of legend windows are available the multiclass legend window and the continuous colour attribute legend window 7 8 1 Multiclass legend window This legend window is displayed when multiclassification is active This occurs when the user does manually set a multi class system see section 7 2 2 or when a fuzzy classifi cation tool is used see section 8 3 3 Objects ar coulored according to the class they belong at most Due to Explorer3D current evolutions additional actions proposed by Explorer3D User s manual 25 this legend window are not currently fully functional nevertheless the user can visualize the centroid of each group appears as a small cross coloured as the group Class dif fusion through objects can also be observed using
5. and waits for text commands sent using the json format The list of commands available is quite simple for the moment and focus on receiving a set of objects features et on setting or updating their class and by the way their colouring e loading objects objects are described by a set of features Their are two parts in the message the command name msg set data then the objects description val object11 object21 Here is an example with 4 objects described by four features vals E Explorer3D User s manual 43 0 09995818 0 8450656 0 31611204 0 9885413 ds L 0 20870173 0 6817836 0 30197513 0 20966005 ls L 0 95933825 0 7643002 0 23336774 0 7409621 Iy L 0 075930595 0 12991834 0 6995411 0 092056274 Js msg set data e objects class definition full set The class of all objects is given the message consists of three parts the command name msg set classes then the number of classes nbclasses value and last the class value for each object sorted in the same order as objects in the set data message val classe_object1 c Here is an example with our set of 4 objects there are 2 classes the 3 first objects belong to class 1 the fourth one belongs to class 0 yal E 1 1 1 0 J msg set classes nbclasses 2 e objects class definition subset this
6. A pie chart chart gives the amount of original variance that 28 M Exbrayat L Martin the current scene offers If a class attribute has been selected then a class purity index is displayed at the bottom of the sub window Dunn index Observations a Variance expliqu e 49 5 Indice de Dunn 0 00386 Figure 20 Observations sub window We can notice that many projection dimensions are available The three most significant have been chosen for the 3D display red bars of the histogram and that they support 49 5 of the original variance pie chart A class attribute has been selected the quality of which is displayed according to the Dunn index 7 11 Project save and load The user can save the current project in order to reload it later The current saving strategy manages data sources 3D scenes and the active additional attributes class colours etc To save the current project choose Save project in the Files menu To load a project choose Load project Projects are by default saved with the e3d file extension Technically speaking they consist of text files using the JSON format 8 Main tools 8 1 Multi view visualization Explorer3D is oriented towards a multi scene interactive use This can be used to follow several goals first to open several visualization windows simultaneously and second to observe the behaviour of a set of selected objects in these various wi
7. Arbre de recouvrement mi 2 Nombre de groupes 8 Arbre de recouvrement minimal Attributs compl mentaires Y Contr le plus proche voisin Y Spec Langage Figure 29 Minimal spanning tree et hierarchical classification Spatial reconfiguration This tool is an experimental one the user is encouraged to test This tool can be reached through the Tools Interaction menu It allows to set a list of relative distance con straints in order to ask for objects to be moved closer or far away and to modify the projection dimensions in order to respect these constraints as much as possible while intro ducing as few global distorsion as possible We firts invite the user to try the Comparison tool 8 4 1 Comparison First check the Object A box and then click on any object within the 3D scene A red target is then displayed around this object fig 30 A green target is also displayed that surround the second object the one the distance to object A should be modified By default this second object is the object with global rank 0 Check the Object B box and click on the second object i e the one you really want to move away or nearer to object A T green target is moved in order to surround this object fig 30 Using the horizontal slider set the expected distance to the left objects have to be moved closer to the right they have to be moved away fig 30
8. Click on OK to validate your constraint Click on Run in order to compute the new projection dimensions When computing is over targets get hidden and the projection space gets modified fig 31 Several constraints should be entered and taken into account for a space modification 36 M Exbrayat L Martin o Explorer3D Fichiers Perspectives Outils Contraintes Comparaison Anomalie D placement Axes d observations O ObjetA rouge n 41 41 v inl a ObjetB vert n 98 98 dist vue 3D dist espace d origine Avant 1 6683884 Apres 1 6659153 0 50 100 valid Comp 41 98 lt 0 85087806 Supprimer la contrainte s lectionner Tout Annuler Recalculer Sauvegarder les changements Charger un fichier Taille des donn es 150x 5 Figure 30 Distance constraints The two objects concerned can be identified in the 3D scene object A 41 red target object B 98 green target The constraint asks for the two objects to be moved closer slider in the left most window and the constraint is textually displayed in the bottom left list Comp 41 98 lt 0 85 0 Explorer3D Fichiers Perspectives Outils Contraintes Comparaison Anomalie D placement Axes d observations ObjetA rouge n 41 41 a ObjetB vert n 98 98 dist vue 3D dist espace d origine Avant 1 6683884 Apres 1 6
9. This projection is computed in order to respect to some end the original spatial distribution e R discriminant Analysis similar to LDA but only the interclass variance is max imized it only focuses on moving away objects of different classes e Rw discriminant analysis similar to r discriminant analysis but the original dis tance is also taken into account the moving away of objects belonging to different classes will be proportional to the inverse of the distance in the original space i e objects that are already far ont from the other in the original feature space will be less affected than ones that were near one from the other The sigma parameter allows to tune the original distance impact The greater sigma the smaller the impact on originally distant objects sigma 1 leads to a r discriminant analysis e KNN PCA experimental not documented e LLE experimental not documented Pre treatments The Projection Method sub window is tightly associated with the Pre treatment sub window This latter is used to apply to data a so called conditioning This window offers various treatments to be done on data before applying the projection method Explorer3D User s manual 17 A check box allows to choose wether data has to be centered or not Centering means that the values of each feature are translated so that their is zero This pre treatment has to be done for ACP and might be suitable for other
10. v Sigma 4 Calculer Noyaux Remise z ro des transformations Donn es repr sent es Observations Courbes 2D Attributs compl mentaires Explorateur 0 L gende Visualisation Infos mod rm f d E Attribut Valeur Classe Ellipsoide Enveloppe convexe Couleur Nom Card EJAcP u E3 0 1 2 C Visualisation 3D lx ES Iris setosa 50 y v LJ r Iris versicolor 50 q 0 1 1 Visualisation 3D Ww LJ Ex Iris virginica 50 GJ ARD 0 1 2 C Visualisation 3D Ej Attribut Valeur Gace E 0 1 2 y Visualisation 35 K il IN Arbre de recouvrement minimal Taille des donn es 150x 5 ACP Diagonalisation OK Figure 21 Several simultaneous views The active view bottom right can be identi fied by its red border We can notice that global controls have been used on the views colouring according to classes as well as local controls the lightgrey background of the upper left scene 8 1 3 View browser The set of available views can be browsed in the View Exlorer sub window This sub window contains a tree view of the current state of explorer3D with data input files data sources as roots followed by the projection method the projection axes and finally the view usually the 3D scene The currently active view is highlighted in this browser Clicking on a given view in the browser gives makes it active 30 M Exbrayat L Martin Figure 22 present
11. 5 nearest neighbors 8 6 SVM SVM stands for support vector machine which are a classification tool that try to separate two groups of objects using a hyper plane They are thus usable when the user knows the class of at least a subset of the objects In the easiest configuration the separation is a plane in the original or projection space that perfectly separates the groups Moreover this plane must be as far as possible of the two groups i e as central as possible Such a clear separation is not always available Either the groups might be somehow mixed or separated by a surface that does not look like a plane For this reason SVM are usually used together with kernels see section 8 7 If more that two groups exist it is possible either to define a SVM between two of the groups or between one group and all the remaining ones In Explorer3D the SVM control are accessed through the SVM perspective The class attribute must have been set before using SVMs Explorer3D User s manual 39 Figure 34 illustrates a SVM simple case based on the very standard iris dataset We can see that a plane has been drawn in the 3D scence that materializes a SVM computed between the iris setosa group on one side and the remaining groups on the other side c Explorer3D Graphem j 0 x ll Visualisation 3D home exbrayat documents recherche developpement jeux iris csv Fichiers Perspectives Outils Aide SVM 1 Cons
12. ARRE AR OA A 2 NT 203 ot 204 205 Q 206 Q 207 208 Q 209 210 NS El 213 ES 214 Q 215 TITTTTTTTTTrrrrrrrrrrrrrrrrr TTTTTTTTTTTTT 216 C AfficherLignes Afff lt h Taille des donn es 2000 x 217 oO Figure 47 Travel throughout dimensions after the axis modification Index 2D graphs 40 Active view 27 Additional attributes 15 Colouring gradient 17 texture 17 Colouring dimension 25 Continuous attribute legend window 23 Crop box 20 spherical 20 Data conditioning 14 Data import 42 Hierarchical classification 32 Interactive projection anomaly 34 comparison 33 move 39 observation axis 35 Kernels 39 Linear discriminant analysis 14 Magnifier 20 Minimum spanning tree 32 Multiclass legend window 22 Multidimensionnel scaling MDS 15 Objects Cloud centering 24 Observations 25 Options 21 pr traitements 9 Pre treatment 14 Principal Component Analysis 14 Projection 3D coordinates 15 distance matrix gt 3D 15 ND gt 3D 14 Projection axes 24 choose 25 colouring 25 manually mixing 44 Sheppard diagram 40 SVM 36 View browser 27 Voronoi cells 40 Zoom on less discriminative dimensions 24 48
13. Kernels subwindow allow the user to select a given kernel in a drop down list and to set the potential parameters of that kernel Once a kernel has been choosen it is applied on any new peojection click on the Compute button of the Projection Methode subwindow to apply the kernel to the current scene M thode de projection ACP Dual automatique Dual forc Calculer Noy aux Polynomiale lv Figure 37 kernel control subwindow We can see that a polynomial kernel has been choosen with a degree 2 polynom We just list the kernels available in Explorer3D and encourage the reader to refer to kernels litterature to get more details e linear kernel this is dummy kernel that allows the user to keep using the original dimension space while kernels are active e polynomial kernel e RBF kernel e Sigmoid e lsomap e LLE Even not kowing much on their specificities the newby user might test them looking for an interesting projection space A2 M Exbrayat L Martin 8 8 2D Graphs The 2D Curves subwindow allow the user to use various additional curves e Shepard diagram this diagram highlights the link between distances in the original and projection spaces e 2D symboles and 2D labels produce the same projection as a 3D scene using only the two first axes The Symboles mode displays each object using the shape of its class The labels mode display the objects labels by de
14. al Iris setosa 50 y Dual automatique C Dual forc Calculer Noyaux Remise z ro des transformations Donn es repr sent es Observations Courbes 2D Attributs compl mentaires Explorateur Ea Iris versicolor 50 l y Es Iris virginica 50 Attribut Valeur ES ACP 3 0 1 2 C Visualisation Taille des donn es 50x 5 Figure 6 Convex envelops and is optimal for that subset according to the projection criteria The crop dialog is described in section 7 6 Interaction A tool to modify the projection axes In this tool one set a list of spatial anomalies and ask for a modified projection that corrects these anomalies See section 8 4 Travel thrgough dimensions A tool to manually set the projection axes by chossing or mixing the axes available See section 9 1 Explore from images A tool to choose which objects to view in the projec tion space together with the associated pictures and nearest neighbors See section 8 5 Options To set various options most of them corresponding to visual ele ments See section 7 7 e Help On line manual Opens a browser the browse the on line version of this manual About Explorer3D Version facts mains authors etc 16 M Exbrayat L Martin 7 Main functions 7 1 Projection methods 7 1 1 Based on a set of features Projection method Several techniques are currently
15. axis In order to observe a subset of objects along a given axis the user can set observations axes For this purpose the user chooses the Observation Axis tab fig 32 and then chooses to objects to from the ends of the axis by clicking on them in the 3D scene The number of objects to be displayed is then set default 10 and the user clicks on OK The 10 objects that are the nearest ones to the segment are then selected and a table containing informatin about these objects global rank and additional attributes is then displayed Objects description window Several axes can be available simultaneously The user selects the ones to be displayed hidden using the Axis window Fichiers Perspectives Outils Contraintes z y TEEN aT Kg Comparaison Anomalie D placement Axes d observations r ObjetA rouge a ObjetB vert Nombre de points a afficher fio PRESS Comp 41 98 lt 0 85087806 Taille des donn es 150x5 Figure 32 Displaying an observation axis we observe the ten nearest objects to axis 8 14 8 5 Explore from Images In order to make the projection space more readable we add the possibility to only display a subset of the objects based on the associated pictures For this tool to work properly the associated data additional attribute must have been set The
16. colouring is automatically used if the color attribute is of numerical type Note Explorer3D does by default consider that additional atributes are of symbolic type For an additional attribute to be considered numerical it must be explicitely declared as in the source file its name being suffixed with N Colours can be replaced with black and white textures The use of textures is activated by checking the Texture box in the Additional attributes subwindow There are cur rently ab ten textures available The user can change the texture associated with a set of objects by clicking on it in the legend window Textures are also displayed on ellipsoids and convex envelopes if these laters are displayed Colours remain displayed in the current release to highlight selected objects a selected object will be coloured even if textures are active it be become textured again if dese lected 20 M Exbrayat L Martin Figure 10 illustrates the use of textures We can see that the user is currently changing the texture associated with the third group iris virginica Visualisation Infos mod Classe Ellipsoide Enveloppe convexe Couleur Nom Card o Choix de texture Texture actuelle ES d rr Iris setosa 50 HNZ LES N l x m Iris versicolor 50 Iris virginica 50 Annuler Mn Uy Mh y is V QS o L gende Axe 12 95008 22 14167 Figure 9 Gradient c
17. method The data variance can be modified depending on the value chosen on the Reduction method list with Raw data variance is not modified with Reduced variables values are rescaled so that the variance gets equal to 1 this option is suitable for PCA with Normalized distances the values are rescaled so that the amplitude gets equal to 2 and thus together with data centering values are between 1 and 1 7 1 2 Based on a distance matrix We currently propose a single projection method which consists in a linear multidimen sional scaling MDS that is a linear projection that computes 3D coordinates so that the distances between pairs of projected objects do correspond as much to their distances in the matrix Theoretically speaking this projection stands if the input matrix does really contain distances with triangular inequality in which case we can consider that these distances were computed in a high dimension space and we thus make some kind of dimensionality reducation No pre treatment is available We did recently had kernel methods Those are presented in a specific section 8 7 Kernel methods can be found in the kernel sub window They allow for the computation of non euclidean distances For instance using the isomap kernal we will obtain a new projection were the only original distances that are kept in the projection space are the ones between nearby objects this kind of kernel is usually s
18. out of 3D window pictures will be displayed either in a separate window or in the 3D view e Viewing pictures dynamically Pictures can get displayed dynamically when the mouse pointer passes over the corresponding object and hidden when the mouse leaves the object This function is enable when the Show pictures when mouse is over option gets selected e Managing options Get in the Tools options menu Explorer3D User s manual 5 e Picture driven exploring You can alternatively choose to hide all ojects in the 3d view and view only the ones that correspond to specific pictures This feature gets activated when you choose Tools Explorer from Images e Moving points closer or away in the 3D frame Exploring from images allox for projection modifications i e to modify the spatial positionning of objects There exists a second mean to achieve projection modification moe complete and also more complex relying on another interface This can be reach trough the Tools interaction menu 4 Install 4 1 Prerequisites Before using Explorer3D you should have installed a recent version of Java Standard Edition gt 1 6u12 along with the Java3D library JavaSE and Java3D can respectively be dowloaded at e http www oracle com technetwork java javase downloads index html e http www oracle com technetwork java javase tech index jsp 138252 html In case java3D is misinstalled or missing a mess
19. proposed in the Projection Method sub window reached through the Perspectives ND gt 3D menu The actual method is chosen from a list Currently available methods are e PCA is the default method it projects objects so that the distance between them in the projection space corresponds as muh as possible to their distance in the original space input features are seen as coordinates in a high dimension space Two options are proposed with PCA dual auto and dual forced Roughly speaking the PCA dual method leads to the same projection as PCA but can be much faster when the number of input features is larger thet the number of objects The user must notice that dual projection forbids the future use of some complementary tools such as interactive projection dual auto compares the number of features and the number of objects and uses the dual method if needed dual forced uses the dual method whatever e LDA This method and all of the following ones rely on the availability of a class attribute defined for each object In other words this supposes that an additional attribute is available that indicates the class of each object and has been explicitely set as the Groups attribute in the Additional attributes sub window see section 7 2 1 Projection is computed so that objects that belong to the same class tend to be grouped while objects that belong to different classes tend to be moved away one from the other
20. several are displayed Magnifying parameters diameter coefficient can be adpted in the Magnifier sub window which is displayed when the Aspect perspective is chosen Figure 13 illustrates the use of the magnifier Visualisation 3D Attribut Valeur iris cvs Figure 13 Magnifier In the left most scene selected objects are one that are magnified to be more precise the ones that are under the magnifier in the right most scene One can notice the magnifier parameters sub window in the control window 7 6 Crop Crop consists in choosing a subset of the objects displayed and then to compute a new projection and scene that fits this subset The resulting projection can be quite different from the original one There are two ways to crop e Box crop the selection zone is a box Such a crop is initiated from the Tools crop menu which displays a crop dialog Clicking on New Crop is this dialog a cubic box gets displayed that is centered at the origin of the scene This box can be moved and or resized using the sliders in the crop dialog fig 14 S If objects were selected when New crop is clicked then the box is automatically sized and moved to contain these objects When the bow has resized and moved as suited clicking on Crop causes a new dataset to be produced that contains the croped object and a new projection and 3D scene t
21. the object to the separating hyperplan see section 8 6 and materialization of a fourth projection axis by the mean of colouring see section 7 9 This legend window does only contain the gradation together the min and max values of the attribute 26 M Exbrayat L Martin 0 L gende Axe PRAXIS 1 MS 2 7741 71 3 3091402 Figure 18 Continuous attribute legend On can notice that the colouring corresponds to a fourth axis materialization that in fact is the same as xl 7 8 3 Texture based legend As introduced in section 7 2 5 colours can be replaced by textures The legend will allow the user to identify and possibly modify the texture associated with a given group 7 9 Projection axes The Projection axes sub window allow several actions on the 3D scene The two first proposed actions are e Zoom on less discriminative dimensions depending on the projection tech nique it might frequently occur that objects get well spread along xl but more concentrated along x2 and even more on x3 To get a better dispersion in the scene the user might check this optionin order to dilate object coordinate on the less exploited axes see figure 19 e Center cloud this option translates the objects so that the origin does not correspond to the baryceeter any more but to the median value along each axis This two options do apply to the current scene and to the ones that are computed after wards Expl
22. u V Y gt 12 Q U 13 V U 14 V Us 15 V Us 16 V Us 17 V Un 18 J Q VC Q gt 20 Y U a Y V 2 Y U 23 V V 24 V U 23 U U 26 V U TOTO O PEP AO ATT rrr ret Y Y Y 91 Ce Ce Taille des donn es 2000 x 217 Figure 45 Travel throughout dimensions objects are projected on the three main projec tion dimensions Explorer3D User s manual 47 295 0 Explorer3D x oO Fichiers Perspectives Outils Coordonn es dans l espace de visualis A X1 X2 X3 1 YU0 y ne x S ee De DL PO a On eae ore a ae gt gt JR ele wan nn 3 yt A Tian EN 9 0 CD on LS se ey YO a v a X j awe IPPO PO OP OO AO TT TT TITO PREP RPP EP PPP ETP PP PTT rrr Arh e as y wf D AT RU on Fu LT w qler A gd a TAG y AS yn a ne gt RAT te GAS b ps ACL LL ML ea EE 19 Q EIA A PEPE br oe n E A ANN a A A tem OR tne Cs gt e qn 4 a COPA eed Oe m TO dt ere PIES A PEPE Mn C SORT SY DIT RE z 22 yA ey y A AD e ZAS AE PT les 26 Q 27 e e Taille des donn es 2000 x 217 Figure 46 Travel throughout dimensions we choose to replace the third dimension with the fourth on axis x3 o Explorer3D x Cl eal Fichiers Perspectives 190 4 E o 191 192 Q 193 Q 194 Q 195 Q 196 Q 197 198 199 Q 200 Q 201 Q 202 TITTI TTITTTTTTT mn
23. user then activates the tool by checking Explore from Images in the Tools menu All objects are then removed hidden from the 3D scene and a new window Associated pictures opens where the user can choose the objects he wants to display fig 33 The user then chooses the pictures he wants to observe either in the drop down list available at the top of the window List tab validate or by typing the picture file name Name tab validate 38 M Exbrayat L Martin Each time a new picture is choosen this picture gets displayed in the window and the corresponding objects is set visible back in the 3D scene When a picture gets clicked the object is swaped to the selected state and the corresponding shape in the 3D scene becomes haighlighted Clicking again on a picture causes the object to be deselected The object n neighbors can be displayed together with their picture the value of n is set by the user in the Neighbors tab and is by default set to 0 When an object gets hidden back so are its neighbors This explroation can be superimposed with the standard dynamic display of pictures in the 3D scene Images associ es RNE isualisation 3D media My Passport Graphem Gamme _Logicielle_ Graphem Resultats Guillaume 2009 11 02 Guillaume1pctL10 cs _ j 0 x Options lt lt From 0 To 39 gt gt Figure 33 Explore from images with
24. 17 7 2 Using displaying additional attributes 17 7 2 1 General Overview 17 7 2 2 Multi Class Objects 18 7 2 3 Image display 18 7 2 4 Multiple picture sources 18 7 2 5 Management of colors 19 7 3 Mouse control 20 7 3 1 Manipulating the 3D scene aoaaa 20 7 3 2 Acting on objects 20 7 3 3 Popup Menu 20 7 4 Object selection 4 21 1 4 1 Multiple selection 21 7 5 Magnifier 22 7 6 Crop 44 4444444 ee ee ei 22 7 7 Options 4444444444 eee 23 7 8 Legend 4 24 7 8 1 Multiclass legend window 24 7 8 2 Continuous attribute legend window 25 7 8 3 Texture based legend 26 7 9 Projection axes 1 26 7 10 Observations 27 7 11 Project save and load 28 8 Main tools 28 8 1 Multi view visualization ee 28 8 1 1 Multi view principle 28 8 1 2 How controls affect 3D scenes 0 0048 29 8 1 3 View browser 29 8 1 4 Selected objects in views 30 8 2 Tabular view of objects 30 8 3 Clustering
25. 31 11100 111 55500 0 01050 0 00009 0 00544 16 67600 137 87100 111 36600 0 00997 0 00009 0 00502 16 01400 141 78100 110 65500 0 01284 0 00011 0 00655 120 55200 131 16200 113 78700 0 00968 0 00008 0 00463 _ i 7 27 2293 aj Es 4 i di gt Figure 43 Import end of import Figure 44 Projection after import Parkinson dataset 46 M Exbrayat L Martin 9 Secondary tools 9 1 Travel throughout dimensions This tool can be reached through the Tools Travel throughout dimensions menu It allows the user to manually set the three projection dimensions by mixing the projection dimensions computed by the system ND gt 3D mode The content of each axis is set by the mean of sliders each slider corresponding to an available projection dimensions Clicking the Draw button causes the objects to move smoothly to their new coordinates see figures 45 to 47 The most simple use of this tool is quite similar to the one of fered by the Projection axes subwindow of the 3D gt ND perspective see section 7 9 Nevertheless using this secondary tool the user can compute any mix of dimensions for instance summing the values of several dimensions on the x3 axis x O Explorer3D Visualisation 3D Attribut Valeur mfeat fac csv Fichiers Perspectives Outils Coordonn es dans l espace de visualis 2 X1 X2 X3 DO N a u W N al YY 8 V V 9 U V 10 Q U
26. 3D view then opens and displays objects using their default shape and colour that is blue spheres on a dark background fig 3 Let us now display an additional visual information by the mean of object colouring according to their class each sphere correspond to a given kind of iris and each kind of 12 M Exbrayat L Martin Courbes 2D Attributs compl mentaires Explorateur a z 9 Attribut Valeur Gq ACP mal 0 1 2 Visualisation Figure 3 Default 3D view iris belongs to one of three varieties The class of a given iris i e its variety is stored in an additional attribute load together with the features and named classe Let us display the sub window Additional attributes by clicking on the upper rigth arrows of this sub window and let us choose classe in the Groups list which is the only attribute available let us ignore multigroups for the moment spheres are then coloured and a legend window pops up fig 4 We can notice the presence of check boxes in the legend window The class check boxes allow to display or hide the objects that belong to a given class The Ellipsoid checkboxes allow to display an ellipsoid around the corresponding group of objects i e the objects of a given class that reflects their spreading in space fig 5 Basically this ellipsoid is centered on the center of gravity of the group and its dia
27. 659153 0 50 100 set cine Valider Comp 41 98 lt 0 85087806 Supprimer la contrainte s lectionner Tout Annuler Recalculer Sauvegarder les changements Charger un fichier Taille des donn es 150x5 Figure 31 Distance constraints the constraint has been integrated in the projection framework We can see that projection has been modified and that the two objects are now closer 8 4 2 Anomaly The Anomaly tool allow the user to modify the relative positionning of three objects We thus manipulate three objects B and C The underlying idea consists in modifying distance A C with regards to distance A B In other words the user expect C to be moved close to or away from A so that in becomes closer or more far away from A than B We proceed as before using the checkboxes and the scene to set B and C and then modifying the distance rate using the slider Explorer3D User s manual 37 8 4 3 Move The underlying idea consists in changing the neighborhood of an object This is a rather complex method the user might use with care The user must first set the target neigh borhood by selecting the objects that belong to it and then by clicking on Add She must then click on object to move and click on the corresponding object Just like before the user than has to click on OK and then on Run 8 4 4 Observation
28. DVP Flo Na 74 99700 d tape 5 Validation du nombre d objets et de descripteurs Votre fichier comporte 195objets d crits par 23 attributs de description et 1 attributs compl mentaires Vous avez fourni les noms des descripteurs Log Nouvelle version MDVP Fo Hz MDVP Fhi Hz MDVI h 9 3021 Figure 42 Import summary 0 Visualisation 3D Attribut Valeur parkinson csv o Outil d importation de donn es Fichiers Version de base e 2 3 4 EF l name ____ MDVP Fo MDVP Fhi MDVP Flo Ma 2 phon RO 119 99200 74 99700 d 3 phon_ RO 122 40000 113 81900 d tape 6 Fin de l import a phon_ RO 116 68200 131 11100 TERN L import est termin 5 phon_RO 116 67600 137 87100 111 36600 Choisissez l une des actions possibles G phon RO 116 01400 141 78100 110 65500 d 7 phon RO 120 55200 131 16200 113 78700 SALES a Charger CE phon RO 120 26700 137 24400 114 82000 d 9 phon RO 107 33200 113 84000 104 31500 d 10 phon_RO 95 73000 132 06800 91 75400 d fa n orozco liaoannannalaLnanacana ld 4 il 3 gt Nouvelle version 195 ES 23 MDVP Fo Hz MDVP Fhi Hz MDVP Flo Hz MDVP Jitter MDVP 19 99200 157 30200 74 99700 0 00784 0 00007 0 00370 22 40000 148 65000 113 81900 0 00968 0 00008 0 00465 116 68200 1
29. Enveloppe convexe Couleur Nom Card a Iris setosa 50 En Iris versicolor 50 ES Iris virginica 50 Figure 5 Groups Ellipsoids list of previously loaded files once a file has been succesfully loaded it gets memorized by Explorer3D and proposed to the user and the bottom of the Files menu Exit To shut down Explorer3D cleanly e Perspectives This menu leads to some relevant sets of sub windows with respect to a given kind data file or to a specific kind of task ND gt 3D Perspective that corresponds to features data files projection from N to 3 dimensions Distance gt 3D Perspective that corresponds to distance matrix files Aspect Tuning of visual parameters e g background colour size of objects etc Classification Offers a set of clustering algorithms e g kmeans gaussian mixtures SVM Offers various SVM visualisation tools that is to says means to visualize a separating hyperplan between two groups of objects see section 8 6 e Tools Crop Data cropping based on a 3D box This consisst in a way to select a subset of objects and compute a new projection that limits to this set of objects Explorer3D User s manual 15 x DO Exp Fichiers Perspectives Outils Pr traitements Param tres ACP Visualisation Infos mod Classe Ellipsoide Enveloppe convexe Couleur Nom Card ACP J
30. Explorer3D User s Manual Release 1 2 Matthieu EXBRAYAT Lionel MARTIN LIFO Universit d Orl ans January 3 2013 Contents 1 2 3 A Foreword on the english documentation What is Explorer3D Explorer3D in a nutshell Install 4 1 Prerequisites oaoa a a a 42 Install and start Data input format 5 1 lt A Foreword on Additional Attributes 5 2 Multisource mode 5 3 Subset files 0000 0 0000 00000000000 000008 5 4 Features input file 5 4 1 General toma 5 4 2 User defined delimiter 5 4 3 Subset data file 0000000000000 5 4 4 How symbolic attributes are managed 5 5 Distance files 4 4 0 000008 5 6 Raw 3D coordinates fle 000000000008 5 6 1 Subset datafile 00 0 0000 000000000000 5 7 Pure additional attributes files a aaa a 5 8 Reserved attribute names additional attributes 9 9 Data import 4 4 Visualisation interface an Overview 6 1 Handover 2 ee ee ee ee 6 2 Control window menus 0 0 0 4 4 4 2 M Exbrayat L Martin 7 Main functions 16 7 1 Projection methods 2 16 7 1 1 Based on a set of features 16 7 1 2 Based on a distance matrix 17 7 1 3 Based on 3D coordinates
31. age should warn you at launch time or at your first try to display a 3D view Meanwhile if no control window get displayed at launch time or misdisplayed window might be the result or the absence wrong use or wrong version of Java or Java3D 4 2 Install and start Explorer3D is shipped as a java archive file explorer3d jar Launching is thus done typing java jar explorer3d jar Depending on your O S e g MacOS you might be able to launch Explorer3D by clicking on the explorer3d jar file Nevertheless this might result in a faulty luanching with no message displayed Double click launch is discouraged with Windows systems An alternative consists in creating a explorer bat file in the directory where explorer3D jar 1s stored and fill this file with java jar explorer3d jar Double clicking on explorer bat should cause Explorer3D to start This alternative start mode is also encouraged in the case of a personnalized install of Java3D where an explicit classpath is needed As the jar option prohibits the use of cp the solution consists in first adding explorer3D jar to the classpath and then directly select explorer Explorer3D as the java class to be started T he following script illustrates how to proceed example with bash et 64 bits java3D library export LD_LIBRARY_PATH LD_LIBRARY_PATH j3d_install_path 1lib amd64 export CLASSPATH CLASSPATH j3d_install_path lib ext j3dcore jar j3d_install_
32. command allows to change the class of some given objects and can only be invoked after the global class setting command has 44 M Exbrayat L Martin been invoked at least once The message consists of two parts the command name msg set classes fast and the list of changes This latter consists of a table val classe_object1 classe_object2 jump_objects the encoding of which is as follows values ar sorted as the objects appeared in the set data command A positive value stands for a new class value of the corre sponding object a negative value indicates a jump in the set objects the class of which is not modified Here is an example with our 4 objects The class of the two first objects is not modified jump 2 The class of the third one is set to 0 The class of the fourth one is not modified jump 1 val s2 0 1 1 msg set classes fast 8 10 Data import The data import tool is open when the user tries to laod a file with an unknown format It might also be directly open to format an input file through the Files Data import tool menu This rather intuitive tool does first load the source file fig 38 The user then indicates the delimiters used in the file field separator dot symbol and text separator fig 39 A view of the input file is available during this step In the proposed example we use the Parkinson dataset UC Irvine machine learning repos
33. convex envelops by checking activate convex envelops and moving the bottom slider envelop either dilate or shrink to the objects that belong to the class at a higher level than the one indicated by the slider The ellipsoid column is not currently used Xi 10 Visualisation 3D 0 L gende Visualisation Infos mod Classe Enveloppe convexe Centre Couleur Nom EM puro rumeanso o BBB autor meansi o Rs auto FkMEANSs2 E AUTO FKMEANS3 g AUTO FkMEANS4a fs E R R El Activer les enveloppes convexes WE Figure 17 Multiclass legend window Here we observe a fuzzy kmeans Each objetct is coloured according to 1ts main class 7 8 2 Continuous attribute legend window The user can define a colouring that does not rely on a given set of class es but rather on a single numerical floating point attribute A colour gradation is then computed to handle any value of the chosen attrifute from the smallest one to the largest one see figure 18 Each object gets coloured depending on the attribute value it is associated with Such a colouring technique is not chosen through the additional attributes sub window This colouring technique can be activate by three kinds of actions selection of an explicitely numerical additional attribute as colouring attribute SVM visualization using colour to materialize the distance of
34. current object rank in the whole set starting at value 0 If the ranks given do not start at 0 then the user must use the 8 M Exbrayat L Martin optional parameter START WITH followed by the offset For instance if his personnal object numbering starts at 1 then he will write START WITH 1 For instance SUBSET START WITH 1 3 4 A B C D classe 55 1 3 5 1 4 0 2 Iris setosa 20 4 9 3 0 1 4 0 2 Iris setosa 110 4 7 3 2 1 3 0 2 Iris setosa The user only gives 3 objects with she has numbered 5 20 and 110 in her own numbering system that starts at 1 These objects are thus numbered 4 19 and 109 in Explorer3D 5 4 4 How symbolic attributes are managed Original attributes are by default supposed to be numerical real Symbolic values can also be handeld Symbolic means that the values do not belong a continuous domain Thus both integers and strings might be considered as symbolic attributes integer will be by default considered as numerical values unless they are be explicitly considered as symbols Symbolic attributes will be denoted by a S suffix in their attribute name Let us observe the following example 151 5 R1 5 R2 5 R3 5 R4 5 R5 Clas A Green YES 19 TRUE B Red NO 17 TRUE C Blue NO 49 FALSE The first four attributes are symbolic Concerning numerical attributes their name can be suffixed by N but this does remain optional In the current implementation of Explorer3D a pre p
35. d to a pre defined visual behaviour of Explorer3D i e automatic object coloring direct link to picture files that should normally be set by hand see section 7 2 1 two names are currently reserved e ImgFileName the column contains picture file names The path to files is relative to the data file directory e g concerning the data file in section 5 5 picture files are supposed to be in the same directory as this data file An absolute path can be set starting from the root directory e g data images enluminures img0001 png Pictures will be displayed either statically on demand on dynamically when the pointer passes over the corresponding object e Class a predefined class for each object In Explorer3D object class is displayed by colouring objects A class attribute can be either numerical or symbolic The class to colour mapping is automatically set Explorer3D does nevertheless allow the user to modify this mapping by hand 5 9 Data import If the data file inner format does not match Explorer3D format a data import tool can be executed This latter will be automatically loaded in case the user tries to load a misformatted file It can also be manually launched choosing the Files Data import Explorer3D User s manual 11 tool menu This wizard tool does allow to re organize rows and columns and guides the user step by step One must notice that only text no binary files feature input files neit
36. e Activate the magnifyer see section 7 5 If the menu is poped up while no object is pointed only global actions are proposed global show hide actions and planar selection 7 4 Object selection a uniform selection mecanism has been implemented There exist several ways to get a object selected from the 3D scene from the separated image windows etc When an object gets selected in one of these views in is also displayed as selected in the other ones For instance when an object gets selected in the 3D scene it gets highlighted in this scene but the corresponding picture if displayed does also get displayed on a red background 7 4 1 Multiple selection Several objects can be selected at once Choosing the planar selection in the popup menu allow the user to draw a rectangle in the 3D view clicking on the background a first time to set the upper left angle and a second time to set the lower right angle of the rectangle see fig 11 All the objects within the rectangle get selected see fig 12 NN Figure 11 Multi selection ongoing Figure 12 Multi slection finished 22 M Exbrayat L Martin 7 5 Magnifier This tool which can be activated through the popup menu displays a leans in the 3D scene that magnifies the objects under it A very interesting side effect comes from the fact that the magnified objects do also get selected The user can thus see how objects of a given view spread in other viws if
37. fault their rank e Contributions e Square cos e Voronoi cells These tools are supposed to help interpreting the 3D scene they rely on a snapshot of the scene in which the Voronoi cell around each object is computed We remind to the user that Voronoi celle in the portion of space that is nearer to a given object point in the space than to any other object The cells coloring depends on the kind of visualisation Proximity tells how far objects are in the original space from a given object The user must select a single object in the 3D scene before opening this kind of graph Distorsion tells how much the distance between two neighbors has been affected by the 3D projection Distorsions do not rely on Voronoi cells but on Delaunay triangle which are complementary to Voronoi cells Neighbors in the 2D graph are linked by a quadrilateral the colour of which runs from white no distorsion to black strong distorsion Proximity and distorsion graphs can be used to compare 2D and 3D projections 2D projections and ND original space 3D projections and ND These tools might not be fully functional currently 8 9 Network driven control This tool allow the user to connect a third tier software to Explorer3D in order to visualize objects and additional data coming from this third tier software This function gets activated through the option window see section 7 7 Explorer3D then listens on a given port currently port 50 000
38. g set of stables is displayed Stables are currently computed for kmeans only 8 3 3 Fuzzy kmeans This method is derived from the former one where each object now belongs to all of the groups at a certain level degree The degree is related to the distanceto the center The sum of degrees is equal to 1 Here again the user can choose the number of centers default 5 On one hand the interest behind this method is to leave more freedom of interpretation to the user On the other hand visualization is bit more tricky We thus developed a specific technic based on convex envelops that we did already partly present in section 7 8 Two kinds of information has to be displayed first the main group of an object and second the set of objects that belong to a group to a given degree The main class is displayed using a classical legend Concerning the degree we use convex envelops see fig 27 for each group or a set of groups we display the set of objects that belong to it to a certain degree i e for which the degree of belonging is higher that x percent remind that we assume that the sum of degrees is 1 The current value of the x threshold can be set using a slider and the evolution of envelops can thus be observed dynamically One can notice that this tool can also be used when a fuzzy classification is given as an input and thus coming from a third tier tool Figure 27 Fuzzy kmeans visualizatio
39. han 4 different values e Colour One different colour is associated with each value of the attribute There is no limit to the number of colours Colouring can not be set directly but can be through the setting of the class attribute 7 2 2 Multi Class Objects When identifying groups of objects the user can either associate a single group to each object or choose a fuzzy classification where each object belongs to each group to a certain level While standard classification relies on a single attribute multiclassification relies on several numerical attributes each one corresponding to the level of belonging to a given class In Explorer3D for a given object the sum of these levels must be equals to Multiclassification can be visualized as follows In the classes dropbox the user chooses the special value Multiclasses A secondary list is then displayed in which the user has to select the list of class attributes The resulting display mode is similar to the one describe in the fuzzy kmean method see section 8 3 3 7 2 3 Image display If an associated data attribute has been defined then the image corresponding to each object can be displayed on the basis of the file name stored in this attribute Image display can be either static or dynamic Dynamic display occurs when the mouse pointer passes over a given object The corresponding image is then displayed and hidden back when the mouse pointer leaves the
40. has not been maintained in the recent releases and should not be used e Show source data array to display in a table the values of additional attributes see section 8 2 e Show picture when mouse is over controls the dynamic display of pictures in the 3D scene see section 7 2 1 e Activate stereo if available If the graphical card and screen of the computer support stereoscopic viewing then stereo gets activated 24 M Exbrayat L Martin Size of labels a slider to set the size of labels in 3D scenes Dynamic object size If checked object size is computed depending on the num ber of number to display the larger the number of objects the smaller their indi vidual size If checked then Size of objects is not available Size of objects A slider to manually set the default object size Multisource mode Allow the user to manipulate multi source data sets see 5 2 Network command available If checked then Explorer3D listens on a port by default port 50000 for some commands The protocol and available commands are presented in section 8 9 Port number to manually set the network command port Currently unavailable Legend available If unchecked the legend window is not displayed Legend windows and their functionalities are presented in section 7 8 295 o Options Couleur par d faut MASA Valorisation de la 3 me dimension au d marrage Afficher les axes originaux avant projection
41. her distance matrices nor rax 3D coorindates are currently supported This does nevertheless cover most of needs Do not hesitate to contact the authors in case additional import features would be necessary To import spreadsheet data please do first save your file using the CVS i e text format and then open this file in the Explorer3D import tool 6 Visualisation interface an Overview 6 1 Handover This section consists in a small tutorial on how to use Explorer3D in the context of a PCA When starting Explorer3D the main control window opens cf fig 1 Pr traitements y Dual automatique L Dual forc Calculer Donn es repr sent es Observations Attributs compl mentaires Taille des donn es 150x5 Figure 1 Explorer3D main control window Figure 2 PCA commands at launch time Let us first load the data file In the menu bar let use choose Files Load local file new project and pick the iris csv file The main window aspect changes so that it now offers the functionnalities that are relevant to a feature file also called the ND imensions 3D imensions perspective see fig 2 In order to compute a classical PCA we first check in the Pre treatments sub window that Center variables is checked and that Reduced variables is selected in the Reduc tion method list Let us then click on Compute in the Projection Method sub window The
42. hosen by the mean of the Column visibility additional attributes menu Projection coor dinates can also be displayed using the column visibility projection coordinates menu By default projection coordinates are not displayed Object selection is availale in this table and the set of objects displayed can be limited to the selected ones see fig 23 for this purpose check Filter selection in the Lines visibility menu For more details on the selection mechanism please see section 7 4 8 3 Clustering Several clustering methods are available They can be reached by choosing the Classifica tion perspective see fig 24 in the main window menu Several of the proposed method Explorer3D User s manual 31 Description d taill e des objets Visibilit Lignes Visibilit wallaby Figure 23 Tabular view of additional attributes with filter shared a common weakness they are fully stable they do not always produce exactly the same set of groups For this reason we encourage the user to compute them several times using the recompute button to ensure a good understanding of the methods stability w r t the current dataset let us also underline that the Recompute button should also be pressed if the classification parameters if any such as the number of groups to produce have been modified Fichiers Perspectives Outils Aide Fichiers Perspectives O
43. istance matrix is given if we have 3 objects a b and c the first line will contain distance and distancea and the second line distance On the first line might occur the optional keyword COMPLETE which means that the full square matrix is given According to our example the first line will contain distance a distance y and distance c the second one distance a distance y et distance c etc If additional attributes are given and only their cardinality is given their names are automatically coined as follows Att1 Att2 etc Here is a sample file very first lines of a inter illumination distance matrix 166 1 0694574 1 1302139 1 0019832 1 0004523 1 0656028 0 96607274 1 1858556 image class ms0001_1 jpg ms0001 Mazarine Fr SW D but 12eme ms0001_2 jpg ms0001 Mazarine Fr SW D but 12eme This data set consists of 166 objects For each of them we have two additional attributes the name of an associated image file and the name of the document the illumination comes from the first object belongs to document ms0001 Mazarine Fr SW D but 12eme and the user might find a picture of the illumination in the file ms0001 1 jpg The second illumination comes from the same document and can be viewed in file ms0001 2 jpg The distance between the first and the second object is 1 0694574 the distance between the first and the third object is 1 1302139 etc 5 6 Ra
44. itory During the next step the source file is cut according to the delimiters and presented e Outil d importation de donn es o Outil d importation de donn es Fichiers Quel Symbole de d cimale utilisez vous I 1 Quel s parateur de texte utilisez vous Votre fichier doit tre en format texte cvs fi ouvrez le d abord avec votre _ e 02 puis cliquez sur le bouton ci dessous wales el Figure 38 Import initial window Figure 39 Import setting delimiters as a table fig 40 the user indicates the kind of data that is stored in this file In the current implementation only feature sets files are managed by the import tool The user then indicates whether the features of an object are organized as a row or as a column fig 41 In our example file a row correspond to the description of an object Explorer3D User s manual 45 and the projection relevant features are stored in rows 2 to 196 and in columns 2 to 24 The first column contains the object names this is an additional attribute The first row contains the names of attributes In the current release description features and additional attributes might not be in terlaced i e there is exactly one block of features and at most one block of additional attributes The import then proposes a summary of the imported data fig 42 then 0 Outil d importation de donn es
45. l 19 Figure 7 Out of 3D scene static image Figure 8 In 3D scene static image dis display a dynamically displayed image play The corresponding object is at the mouse over the object upper left corner of the picture achieved with Explorer3D consists in creating several 3D scenes possibly using the same projection and then associating a specific image attribute to each of these scenes For this purpose the user must first activate the view a specific image attribute will be associated with then check the Specific A D box in the Additional attributes sub window and last choose the specific picture attribute Any scene for which the Specific A D box has not been checked will use the global image attribute if any has been defined Conversely several scenes can use a specific image attribute Multiple picture sources can exist for instance in the case multisource data each of this source bringing its own set of images or more generally in the case images are just a side information of the object studied and not the sources of features 7 2 5 Management of colors By default colors are associated with classes which means that setting a class attribute implies that the same attribute is automatically set as the color one Nevertheless it is possible to choose a color attribute directly and thus to dissociate it from the role of class attribute This also allows gradient colouring see fig 9 Gradi ent
46. meters correspond to the variance of the group along its three mains axes based on the hypothesis of a multinormal distribution Figure 5 illustrate a visualisation of objects and the classes ellipsoids One must notice that the top line check boxes blue background are shortcuts to check or uncheck the whole column It is also possible to view the convex envelop of a group This is done by checking boxes in the Convex Envelop column Figure 6 illustrates a 3D view where objects have been hiden and only convex envelops are displayed Last we can interact with the 3D view by using the three mouse buttons left rotation center zoom right translation Explorer3D User s manual 13 0 L gende Visualisation Classe Ellipsoide Enveloppe convexe Couleur Courbes 2D Attributs compl Vv omatique ee a ra mentaires classe Taille des donn es 150x5 6 2 Co Figure 4 3D View with colour and legend ntrol window menus e Files data loading tools Load local file allow to load the various supported kind of data files The file content is automatically discovered by Explorer3D If no standard for mat is detected an error message is displayed and the data import tool is automatically opened Let us remind the three kind of data files managed by Explorer3D x Feature files the file is s
47. modified once the user clicks on Draw This manually set projection does not generate a new scene but rather move objects in the current scene The dimensions available are numbered from 1 to n these are the dimensions computed by the projectin method not the raw input features An additional value is given denoted X which indicates that the corresponding axis is not used For instance setting x3 to X means that objects with all be placed in the x1 x2 plan Last we can choose a projection dimension to colour object continuous attribute colour ing The coulouring dimensions can be chosen among the projection dimensions com puted by the projection method not among the raw features this might proposed in a further release Checking the Colouring dimension box the user activates the continu ous attribute colouring based on the dimension value of each object Figure 18 illustrates the dimension based colouring We can see in this figure that we used the same dimension for colouring as the one used to project objects along x1 Nevertheless any other available projection dimension might have been used 7 10 Observations This sub window indicates the amount of original variance that is displayed using the 3 current projection dimensions figure 20 a histogramm gives the amount of variance each available projection dimension brings currently used projection dimensions are drawn red unused ones are drawn blue
48. n with convex envelops and threshold 8 3 4 Density based classification DBSCAN Differing from the former methods that rely on the distance to a given group center and thus lead to spherically structured groups density based methods build groups starting from the dense parts of the objects space Objects are then aggregated in groups de pending on their neighboorhood diffusion Such methods are relevant when objects are structured in dense arbitrary shaped groups 34 M Exbrayat L Martin Two parameters can be manually set first the neighborhood radius the maximum dis tance within which two objects are considered as neighbors and the minimum group cardinality how many neighbors are needed at least to form a group In order to help the user to set the radius a sphere is displayed at the enter of the 3D scene Figure 28 presents the result of a DBSCAN clustering We can notice that the first group consists of the isolated objects i e the ones that do not belong to a cluster This group might be empty Figure 28 Density based classification dbscan 8 3 5 Minimal spanning tree and hierarchical classification The minimum spanning tree is a tree that links the objects displayed Briefly speaking this tree links each object to its nearest neighbor starting from the nearest pair of objects Visualising this tree shows how objects are linked from near to near A long link will highlight the existence of two distant gro
49. ndows 8 1 1 Multi view principle The term view does mainly correspond to the 3d scene window It also corresponds to less used windows such as tables or 2D graphs The user can compute and manipulate several 3D scenes for a unique dataset mainly to e observe the same projection under several points of view e observe different projections on a unique dataset PCA LDA etc e observe both a global view and one or several crops Explorer3D User s manual 29 8 1 2 How controls affect 3D scenes Some control have a global effect on 3D scenes colouring is currently one of them mainly to identify the of classes in several scenes Conversely many control are not supposed to have a global effect and rather a local one on a single 3D scene this stands for clustering crops etc For this purpose we introduce the notion of active view the active view is the one that is affected by the execution of a non global control action How can we set and highlight the active view A view can be set active by giving the focus to its window Once the view has been set active the window border gets red see figure 21 A view remains active until another one is set active either explicitely or implicitly For instance computing a new 3D scene makes it active Only one view at most is active at any time Fichiers Perspectives Outils Pr traitements M thode de projection Analyse rw discriminante
50. o be computed Croped view can be croped in turn and so on e Spherical crop This crop zone is centered on a given object fig 15 Such a crop is accessed through the 3D scene popup menu Explorer3D User s manual 23 Figure 14 Box crop Figure 15 Sperical crop 7 7 Options The option window is accessed through the Tools options menu The current options are e Default color of the objects The cvoloured rectangle can be clicked in order to modify this default colour e Enhancement of 3rd dimension at display time When the 3D scene gets displayed a slight rotation is applied so that the 3 projections axes get visible e Display original axes Original axes i e input features get displayed in the projection sub space This option affectes only feature based projections This action should moved away from the option list in a further release e Draw pictures out of 3D window to choose whether statically displayed pic tures get displayed in the 3D scene or in a separate window default value e Autocompute 3D projection If not checked the user has to choose and launch a projection technique once data has been loaded Otherwise the projection and 3D scene are automatically computed In case of a feature data source the default projection technique is PCA e Activate distant execution this option should be checked in case the user wants to display the 3D scene on an image wall This old option
51. object Dynamic display of images is optionanl and occurs only if Show pictures when mouse is over is checked Static image display is controled through the popup menu that opens when the user makes a right clic on a given object with her mouse see section 7 3 3 Depending on user s options see section 7 7 image static display will occur either in the 3D view or in a specific window If option Draw pictures out of 3D is checked then a specific window is use see fig 7 otherwise they are directly displayed in the 3D view see fig 8 We can notice in figure 7 that the objected linked to the pictures have been lected red background in the picture window and highlighted color in the 3D view We can also notice that one image is currently dynamically displayed in the 3D view The default option consists in displaying images out of the 3D view in a spcific window 7 2 4 Multiple picture sources If several images are available for a single object the user can decide to display them simultaneously using a set of 3D scenes To be more precise the user can define a specific image attribute for each of her 3D scenes The default configuration consists in using a single image attribute for all of the 3D scenes If several additional attributes are available that correspond to the name of picture files and if the user wishes to view a least two of them simultaneously the way this can be Explorer3D User s manua
52. orer3D User s manual 27 Visualisation 3D Attribut Valeur iris cvs Fichiers Perspectives Outils Pr traitements M thode de projection acp nl v Dual automatique C Dual forc Calculer Noyaux Remise z ro des transformations SE Remettre z ro age AH RH BH Afficher o u 14 Dimension couleur iE Mhcaruatinne Figure 19 Zoom on less discriminative dimensions et cloud centering The left most scene corresponds to a basic PCA projection the right most one to the PCA where the two options have been used The scenes have been rotated so that axes x2 and x3 can be well observed We can see that objects get more spreaded along x3 in the right most scene than in the left most one this can be also observe while less important along x2 and that the object cloud has been translated to the left centering compared to the one in the left most scene A third action is available that alow to choose the projection axes This functionnality does only make sense when at least three axes are available that is to say with a nD gt 3D projection based on feature input data Nevertheless the user can always choose at least to invert the dimension to axes mapping or to use the same dimension along several axes The user can specify the projection dimension used on each axes in the following order x1 breadth x2 heigth and x3 depth The scene is only
53. oulouring Figure 10 Textures 7 3 Mouse control Using a 3 button mouse is recommended Many tasks can be achieved throught he mouse selecting objects displaying information etc 7 3 1 Manipulating the 3D scene Pressing the mouse buttons on the background i e with no object under the mouse pointer the user obtains the following behaviours e Left press move rotation e Center press move zoom e Right button move translation 7 3 2 Acting on objects e Shift left click select deselect the object under the mouse pointer a selected object is highlighted in the scene Object selection is used in several tasks e Right click Opens the popup menu 7 3 3 Popup Menu Right clicking in the 3D scene opens a popup menu form which various actions can be achieved When opening the menu while the mouse points on a projected objects the following ations are proposed Explorer3D User s manual 21 e Show hide the picture associated with the object if a picture attribute has been defined e Show hide all pictures same remark e Show hide the label associated with the object If no label attribute has been defined the object number is displayed e Show hide all labels e Show hide labels for the selected objects e Make a spherical crop centered on the currently pointed object See section 7 6 e Make a planar selection See section 7 4 1 e Center the scene on the currently pointed object
54. path lib ext j3dutils jar j3d_install_path lib ext vecmath jar explorer_install_path explorer3D jar java explorer Explorer3D 6 M Exbrayat L Martin 5 Data input format 5 1 A Foreword on Additional Attributes Besides projection data e g features or distance matrices input files might contain additional attributes handling for instance the name of an associated picture file the class of the object if this stands and so on An input file might contain 0 1 or several such attributes 5 2 Multisource mode In case the user possesses several data source file that concern a single set of objects each of these files containing different features several of these files can be loaded in a single project and projected simultaneously This mode is allowed depending on Explorer3D options see section 7 7 If this option is not activated then each loading of a data source file will result in the reintialization of Explorer3D and thus to discarding the current projection Otherwise the user will be asked if a new file is to be considered to contain a new set of features for the currently observed objects or consists of a new set of objects if which case Explorer3D will be reinitialized In multisource mode each data source file can contain whatever data input format is allowed features distance matrices etc Objects must appear in the same order in each of the files unless the file is a subset one 5 3 Subset files
55. rocessing is done on symbolic at tributes in order to binarize them A first scan of the column is done in orderto list the existing values of the attribute The unique attribute is then replaced by a list of attributes taking value 0 or 1 each one corresponding to a value of the symbolic one For a given object all of these attributes are set to 0 except the one corresponding to the original symbolic value Such forged attributes are named as the original attribute except they are suffixed by Srank where rank is the rank of the symbolic value in the list of existing ones for instance in our example data set R1 get replaced with two attributes R1 1 and R1 2 the values of which are respectively for the first object 1 et 0 and for the second object 0 and 1 This decomposition is fully transparent to the user Explorer3D User s manual 9 5 5 Distance files Distance files contain distance matrices 1 e matrices the elements of which correspond to the distance between pairs of objects The 3D view is then computed so that the distance in the 3D space correspond as much as possible to the distances in the matrix The file inner format is as follows number of objects COMPLETE distances between objects one object per line Number ou list of names of additional attributes Values of additional attributes Where the squares braces denote optional parts By default only the upper triangular part of the d
56. s object spatial positionning in a 2D or 3D space based on various projection techniques e When attribute value files are given Explorer3D relies on dimensionality reduction either supervised e g linear discriminant analysis or unsupervised e g principal component analysis e When a distance file is given distance or dissimilarity between pairs of objects Explorer3D relies on multidimensional scaling Various unsupervised classification tool are available in ordre to help the user s graphical interpretation Explorer3D is globally oriented towards an interactive use in order to allow for subset extractions positionning constraints 3 Explorer3D in a nutshell e Loading a file in menu Files choose load a file If data cannot be directly interpreted by Explorer3D an import tool will open automatically see section 5 9 e Viewing pictures associated to objects In order to use this feature your source file must contain a column that indicates the name of the file containing the picture associated with the corresponding object see section 5 Once your data file has been loaded and visualized go to Additional features sub window and associate Associated data with the relevant column i e the one containing the picture file names e Viewing pictures permanently in the 3D frame righ click on the object and select Show hide picture Depending on option settings Draw pictures
57. s the view browser We can notice that we currently use an attribute value feature set data source on which we computed a PCA and used the 3 main axes of this latter to compute the 3D scene We can also observe that we used another projection method were the 3rd projection axis is set to 1 which means that the x3 z axis is not used Figure 22 View browser An Attribut Value feature set data source has been used on which both a PCA and a LDA have been computed A crop has been computed on one of these views there is a secondary data source that is a child of the first one The crop based scene is currently active grey background 8 1 4 Selected objects in views Selected objects are highlighted in the 3D scenes The set of selected objects is global to all views This means that selecting an object in one view highlights it in all the views This can be usefull to look for a given set of object s throughout several views 8 2 Tabular view of objects Options allow the user to view a table that contains the additional attributes for the current object set As defined previously additional attributes consist of data that are not used to compute the object projection but rather to bring additional knowledge on the objects This table contains all of the objects sorted according to their index number By default all additional attributes are available Whether each attribute is displayed can be c
58. sent es y Observations Courbes 2D y Attributs compl mentaires Y Contr le plus proche voisin Y Taille des donn es 150x 5 Figure 25 Gaussian mixture 8 3 2 Kmeans The principle under that very standard hard clustering method consists in producing clusters of similar diameters To be more precise this algorithm minimizes the sum of the distance of each object to the nearest cluster center The user first sets the expected number of groups Group Centers are then randomly initialized and each object is put in to the group of the nearest center Each center is then updated to the barycenter of the group objects This process is iterated until stabilization or a maximum number of iterations is reached In our current implementation the default number of centers is set to 5 The user can modify this number and recompute the corresponding clustering result Figure 26 Kmeans Kmeans being not stable an additional tool is proposed that computes the stable compo nents This tool is accessible through the stables subwindow a stable is a set of object that are always put in the same cluster by a given clustering method i e kmeans in Explorer3D User s manual 33 our software By always we mean in any of a given number of runs of the clustering method In this tool we can choose the number of classes and also the number of runs to be conducted Once the runs have all been done the resultin
59. tant Choisissez entre la classe Iris setosa v Et le reste v Mode d affichage du SVM a Choisissez un mode de calcul du SVM svm sur les donn es projet es Y S lectionnez une mode d affichage plan v Afficher Attributs compl mentaires Modalit classe v Donn es associ es Label Forme Couleur classe v Noyaux y Taille des donn es 150 x 5 Specif Data Spec Langage Mat Ind Attr Mat Similarite VISU 3D Mat Distance Figure 34 SVM with separating plane iris dataset Availables commands are as follows e SVM subwindow to defined the groups to be separated constant do not modify this value which is currently unused 1 first group one class among the available ones second group either one of the remaining class or all of them e SVM Display Mode subwindow computation mode the SVM can be computed either based on the orginal dimensions input features or on the projection space An original dimension based SVM will be more accurate but it will not be possible to draw it as a plane in the 3D scene display mode the simplest and easiest to interpret display mode consists of a plane Nevertheless two additional modes are available that allow some kind of materialization when plane display is not possible coloring et replace 3rd axis 40 M Exbra
60. tructured like a table where objects are described by a set of numerical or symbolic features 3D projection will be computed by the mean of a dimensionnality reduction method e g PCA x Distance matrix files used when the only available data concern the dis tance or di similarity between objects Projection will be computed by the mean of a MDS Raw 3D coordinates when we directly get such data for instance using the output of another software Load online file this does permit to load a file stored on a distant machine on HTTP server The secondary that opens then allow to set the site URL and then to explorer the web site directories Supported file formats are the same as for local files Data import tool This tool allow to the content of a file text and to port it to a Explorer3D supported file format 14 x O Ex Fichiers Perspectives Outils Pr traitements Param tres ACP ACP _ y Dual automatique C Dual forc Calculer Noyaux Remise z ro des transformations Donn es repr sent es Observations Courbes 2D Attributs compl mentaires M Exbrayat L Martin 0 L gende Classe dasse Donn es associ es Label Forme Couleur Explorateur m CJ Attribut Valeur 9 EJAcCP 43 0 1 2 y Visualisation Taille des donn es 150x5 Visualisation Infos mod Classe Ellipsoide
61. uited when the user knows or guesses that objects are placed on the surface of some non plane surface i e a mathematical variety 7 1 3 Based on 3D coordinates In the case of 3D coordinates projection is identical to input data Centering and dimen sion dilatation are available see section 7 4 7 2 Using displaying additional attributes 7 2 1 General Overview Additional attributes are given as input but not directly used in the object spatial pro jection They can for instance consist of the class of the object an associated image file name etc Five kinds of display modes are currently available e classes The selected attribute consists of the class of the objects By default this attribute is also used to colour objects but colouring is not the only goal of the class attribute Fuzzy classes can also be visualized In this case each object is supposed to belong to each group to a given level the sum of these levels being equal to 1 Fuzzy classes visualization is described in section 7 2 2 e associated data In the current release this function corresponds to the associated images visualization e Label the textual value of the attribute is display in the projection space near the object 18 M Exbrayat L Martin e Shape A specific shape is associated with each value of the attribute Only a limited set of shapes are currently avaible and this should no be used with an attribute taking more t
62. ups etc A hierarchical classification consists in grouping objects starting from the nearest ones always adding the nearest remaining one which might be either a object or an already formed group There exists several ways to compute the distance between two groups The one we use is called minimal jump The distance between an object and a group is the distance between the object and the object opf the group that is the nearest to him Between two groups it is the minimum distance between two objects that belong to each of the group respectively We produce a classification tree called dendrogram Dendrograms are very common for instance to classify species While they are usually display in a flat manner we propose to view them directly in the 3D scene see fig 29 The reader might notice the strong link between the minimum spanning tree and the hierarchical classification based on minimal jump The number of clusters to be formed is set by the user Forming the groups does simply correspond to cut the tree at a certain level starting from the root where the number of branches equals the expected number of groups Explorer3D User s manual 39 8 4 Explorer3D Grapnem 0 x gernle Eme Logicielle Graphem Resultats Guillaume 2009 11 02 Guillaume1pctL10 cs _ o x EA Aide Jf Modalit Ellipsoide Enveloppe convexe Couleur o Classification Choisir une m thode v Recalculer
63. utils Aide Classification a Classification T Choisir une m thode Y DBSCAN kx Recalculer i Recalculer Nombre de points minimum Arbre de recouvrement mi E Nombre de groupes Rayon de la sph re 5 12 Arbre de recouvrement minimal E Afficher Recalculer Stables Arbre de recouvrement 2 Nombre de classes pour KMeans Nombre de groupes 5 5 Nombre de runs de KMeans Arbre de recouvrement minimal 5 Afficher Recalculer Stables Donn es repr sent es z Nombre de classes pour KMea i E Observations Nombre de runs de KMeans Courbes 2D Y E Attributs compl mentaires Y Recalculer Contr le plus proche voisin Y Donn es repr sent es y Figure 24 classification perspective 8 3 1 Gaussian mixture Gaussian mixture classification is built on the idea that objects are distributed according to a ste of multivariate gaussian laws A technique is thus use to compute the most likely set of underlying laws Each computed law is displayed by the mean of an ellipsoid that 32 M Exbrayat L Martin is centered oriented and sized as the law see fig 25 The current method which relies on the klustakwik algorithm does automatically compute the number of laws gt E Fichiers Perspectives Outils Aide Classification a M lange de lois lv Recalculer Arbre de recouvrement min Y Stables y Donn es repr
64. w 3D coordinates file This kind of file does start by the number of objects it contains it is computed automat ically This might be reconsider in some future released of Explorer3D Optionnaly the file might start by a line that contains the names of the 3 attributes Each remaining line contains the 3 coordinates real values of an object These values might be followed by additional attributes Here is a example file with three objects 10 M Exbrayat L Martin 0 5 0 5 0 5 0 5 0 5 0 5 00 2 5 6 1 Subset data file Optionnaly the keyword SUBSET should be added alone on the first line of the file Each remaining line will thus consist of four values the first one beeing the global rank of each object As with feature input files SUBSET might be followed by START WITH to specify an offset for instance SUBSET START WITH 1 40 5 0 5 0 5 2 0 5 0 5 0 5 100 2 This file contains the coordinates of objects number 4 2 and 1 with an offset of 1 that is to say that the global rank of these objects are 3 1 and 0 5 7 Pure additional attributes files the user can load files that consist of additional attributes with no projection data Such files are structured as feature input file where the number of projection features is set to 0 Such files do only make sense in multisource mode 5 8 Reserved attribute names additional attributes Some reserved attribute names might be used that are linke
65. y default the space char acter This delimiter might be explicitly defined and changed in case the space character can not be used see section5 4 2 Here is a sample file first lines of the very standard iris data set 150 4 A B C D class 5 1 3 5 1 4 0 2 Iris setosa 4 9 3 0 1 4 0 2 Iris setosa 4 7 3 2 1 3 0 2 Iris setosa In this sample 4 original attributes are given named A B C and D they are followed by an additional attribute class In this sample the original attributes consist of numerical values reals Remark the user should be careful not to leave empty spaces at the end of lines especially in the two first lines Such empty spaces might be misinter preted by Explorer3D 5 4 2 User defined delimiter If the space character might not be used as a delimiter an additional line can be added at the beginning of the file on its 3rd line thus after the number of attributes and before the names of attributes to declare the delimiter used for instance if is the delimiter the file starts like 150 4 AIBICIDIclass 5 113 511 410 2 Iris setosa 4 913 011 410 21Iris setosa 4 713 211 310 21Iris setosa 5 4 3 Subset data file If SUBSET is set on the first line then the file is supposed to contain only a subset of the global set of objects An additional numeric value must then be added at the begining of each object description line that corresponds to the
66. yat L Martin The colouring mode uses a colour grade to recolor the objects relatively to their side and distance to the hyperplane red on one side green on the other one Figure 35 illustrates this mode the replace 3rd axis used the x3 projection axis to materialize the side and distance to the hyperplane for each object Projection based on input features is then limited to the x1 and x2 axes Figure 36 illustrates this mode Theses two modes can be used with both SVM computed in the original or projected space Figure 35 SVM coloring display mode Figure 36 SVM with 3rd axis display mode Explorer3D User s manual 41 8 7 Kernels Kernels are powerful tools that have been widely used in machin learning due originally to SVM As we said before identified an separating hyperplane is not always easy nor feasible both in the original and projected space Kernels compute new attributes based on some combinations of the original ones Most of these combinations are not linear but must anyway respect some criterias The trick with kernels is that in fact they all allow to compute a distance between object without really computing the new attributes and thus propose various and potentially complex spaces at a low cost We invite the reader to consult the abundant bibliography of the field for more details In Explorer3D kernels are available in the SVM and ND gt 3D perspectives The

User`s Manual Release 1.2

Contents

Download Pdf Manuals

Related Search

Related Contents