Computer Science Technical Report

Size: px
Start display at page:

Download "Computer Science Technical Report"

Transcription

1 Computer Scence echncal Report NLYSIS OF PCSED ND FISHER DISCRIMINNSED IMGE RECOGNIION LGORIHMS Wendy S. Yambor July echncal Report CS3 Computer Scence Department Colorado State Unversty Fort Collns, CO Phone: (97) Fa: (97) WWW:

2 HESIS NLYSIS OF PCSED ND FISHER DISCRIMINNSED IMGE RECOGNIION LGORIHMS Submtted by Wendy S. Yambor Department of Computer Scence In Partal Fulfllment of the Requrements For the Degree of Master of Scence Colorado State Unversty Fort Collns, Colorado Summer

3 COLORDO SE UNIVERSIY July 6, WE HEREY RECOMMEND H HE HESIS PREPRED UNDER OUR SUPERVISION Y WENDY S. YMOR ENILED NLYSIS OF PCSED ND FISHER DISCRIMINNSED IMGE RECOGNIION LGORIHMS E CCEPED S FULFILLING IN PR REQUIREMENS FOR HE DEGREE OF MSER OF SCIENCE. Commttee on Graduate Work dvsor Codvsor Department Head

4 SRC OF HESIS NLYSIS OF PCSED ND FISHER DISCRIMINNSED IMGE RECOGNIION LGORIHMS One method of dentfyng mages s to measure the smlarty between mages. hs s accomplshed by usng measures such as the L norm, L norm, covarance, Mahalanobs dstance, and correlaton. hese smlarty measures can be calculated on the mages n ther orgnal space or on the mages projected nto a new space. I dscuss two alternatve spaces n whch these smlarty measures may be calculated, the subspace created by the egenvectors of the covarance matr of the tranng data and the subspace created by the Fsher bass vectors of the data. Varatons of these spaces wll be dscussed as well as the behavor of smlarty measures wthn these spaces. Eperments are presented comparng recognton rates for dfferent smlarty measures and spaces usng hand labeled magery from two domans: human face recognton and classfyng an mage as a cat or a dog. Wendy S. Yambor Computer Scence Department Colorado State Unversty Fort Collns, CO 853 Summer

5 cknowledgments I thank my commttee, Ross everdge, ruce Draper, Mcheal Krby, and dele Howe, for ther support and knowledge over the past two years. Every member of my commttee has been nvolved n some aspect of ths thess. It s through ther nterest and persuason that I ganed knowledge n ths feld. I thank Jonathon Phllps for provdng me wth the results and mages from the FERE evaluaton. Furthermore, I thank Jonathon for patently answerng numerous questons. v

6 able of Contents. Introducton. Prevous Work.. General lgorthm....3 Why Study hese Subspaces?. 3.4 Organzaton of Followng Sectons 4. Egenspace Projecton 5. Recognzng Images Usng Egenspace, utoral on Orgnal Method utoral for Snapshot Method of Egenspace Projecton Varatons Fsher Dscrmnants 5 3. Fsher Dscrmnants utoral (Orgnal Method) Fsher Dscrmnants utoral (Orthonormal ass Method) Varatons 9 4. Egenvector Selecton Orderng Egenvectors by LkeImage Dfference Smlarty & Dstance Measures re smlarty measures the same nsde and outsde of egenspace? Eperments Datasets he Cat & Dog Dataset he FERE Dataset he Restructured FERE Dataset aggng and Combnng Smlarty Measures ddng Dstance Measures Dstance Measure ggregaton Correlatng Dstance Metrcs LkeImage Dfference on the FERE dataset Cat & Dog Eperments FERE Eperments Concluson Eperment Summary Future Work.. 66 ppend I 68 References 69 v

7 . Introducton wo mage recognton systems are eamned, egenspace projecton and Fsher dscrmnants. Each of these systems eamnes mages n a subspace. he egenvectors of the covarance matr of the tranng data create the egenspace. he bass vectors calculated by Fsher dscrmnants create the Fsher dscrmnants subspace. Varatons of these subspaces are eamned. he frst varaton s the selecton of vectors used to create the subspaces. he second varaton s the measurement used to calculate the dfference between mages projected nto these subspaces. Eperments are performed to test hypotheses regardng the relatve performance of subspace and dfference measures. Nether egenspace projecton nor Fsher dscrmnants are new deas. oth have been eamned by researches for many years. It s the work of these researches that has helped to revolutonze mage recognton and brng face recognton to the pont where t s now usable n ndustry.. Prevous Work Projectng mages nto egenspace s a standard procedure for many appearancebased object recognton algorthms. basc eplanaton of egenspace projecton s provded by []. Mchael Krby was the frst to ntroduce the dea of the lowdmensonal characterzaton of faces. Eamples of hs use of egenspace projecton can be found n [7,8,6]. urk & Pentland worked wth egenspace projecton for face recognton [].

8 More recently Shree Nayar used egenspace projecton to dentfy objects usng a turntable to vew objects at dfferent angles as eplaned n []. R.. Fsher developed Fsher s lnear dscrmnant n the 93 s [5]. Not untl recently have Fsher dscrmnants been utlzed for object recognton. n eplanaton of Fsher dscrmnants can be found n [4]. Swets and Weng used Fsher dscrmnants to cluster mages for the purpose of dentfcaton n 996 [8,9,3]. elhumeur, Hespanha, and Kregman also used Fsher dscrmnants to dentfy faces, by tranng and testng wth several faces under dfferent lghtng [].. General lgorthm n mage may be vewed as a vector of pels where the value of each entry n the vector s the grayscale value of the correspondng pel. For eample, an 88 mage may be unwrapped and treated as a vector of length 64. he mage s sad to st n Ndmensonal space, where N s the number of pels (and the length of the vector). hs vector representaton of the mage s consdered to be the orgnal space of the mage. he orgnal space of an mage s just one of nfntely many spaces n whch the mage can be eamned. wo specfc subspaces are the subspace created by the egenvectors of the covarance matr of the tranng data and the bass vectors calculated by Fsher dscrmnants. he majorty of subspaces, ncludng egenspace, do not optmze dscrmnaton characterstcs. Egenspace optmzes varance among the mages. he

9 ecepton to ths statement s Fsher dscrmnants, whch does optmze dscrmnaton characterstcs. lthough some of the detals may vary, there s a basc algorthm for dentfyng mages by projectng them nto a subspace. Frst one selects a subspace on whch to project the mages. Once ths subspace s selected, all tranng mages are projected nto ths subspace. Net each test mage s projected nto ths subspace. Each test mage s compared to all the tranng mages by a smlarty or dstance measure, the tranng mage found to be most smlar or closest to the test mage s used to dentfy the test mage..3 Why Study hese Subspaces? Projectng mages nto subspaces has been studed for many years as dscussed n the prevous work secton. he research nto these subspaces has helped to revolutonze mage recognton algorthms, specfcally face recognton. When studyng these subspaces an nterestng queston arses: under what condtons does projectng an mage nto a subspace mprove performance. he answer to ths queston s not an easy one. What specfc subspace (f any at all) mproves performance depends on the specfc problem. Furthermore, varatons wthn the subspace also effect performance. For eample, the selecton of vectors to create the subspace and measures to decde whch mages are a closest match, both effect performance. 3

10 .4 Organzaton of Followng Sectons I dscuss two alternatve spaces commonly used to dentfy mages. In chapter, I dscuss egenspaces. Egenspace projecton, also know as KarhunenLoeve (KL) and Prncpal Component nalyss (PC), projects mages nto a subspace such that the frst orthogonal dmenson of ths subspace captures the greatest amount of varance among the mages and the last dmenson of ths subspace captures the least amount of varance among the mages. wo methods of creatng an egenspace are eamned, the orgnal method and a method desgned for hghresoluton mages know as the snapshot method. In chapter 3, Fsher dscrmnants s dscussed. Fsher dscrmnants project mages such that mages of the same class are close to each other whle mages of dfferent classes are far apart. wo methods of calculatng Fsher dscrmnants are eamned. One method s the orgnal method and the other method frst projects the mages nto an orthonormal bass defnng a subspace spanned by the tranng set. Once mages are projected nto one of these spaces, a smlarty measure s used to decde whch mages are closest matches. Chapter 4 dscusses varatons of these two methods, such as methods of selectng specfc egenvectors to create the subspace and smlarty measures. In chapter 5, I dscuss eperments performed on both these methods on two datasets. he frst dataset s the Cat & Dog dataset, whch was developed at Colorado State Unversty. he second dataset s the FERE dataset, whch was made avalable to me by Jonathan Phllps at the Natonal Insttute of Standard and echnology [,,3]. 4

11 . Egenspace Projecton Egenspace s calculated by dentfyng the egenvectors of the covarance matr derved from a set of tranng mages. he egenvectors correspondng to nonzero egenvalues of the covarance matr form an orthonormal bass that rotates and/or reflects the mages n the Ndmensonal space. Specfcally, each mage s stored n a vector of sze N. [ ]... () he mages are mean centered by subtractng the mean mage from each mage vector. N P m, where m P () hese vectors are combned, sdebysde, to create a data matr of sze NP (where P s the number of mages). P [... ] X (3) he data matr X s multpled by ts transpose to calculate the covarance matr. Ω XX (4) hs covarance matr has up to P egenvectors assocated wth nonzero egenvalues, assumng P<N. he egenvectors are sorted, hgh to low, accordng to ther assocated egenvalues. he egenvector assocated wth the largest egenvalue s the egenvector he bar notaton here s slghtly nonstandard, but s ntended to suggest the relatonshp to the mean. complete glossary of symbols appears n ppend I. 5

12 that fnds the greatest varance n the mages. he egenvector assocated wth the second largest egenvalue s the egenvector that fnds the second most varance n the mages. hs trend contnues untl the smallest egenvalue s assocated wth the egenvector that fnds the least varance n the mages.. Recognzng Images Usng Egenspace, utoral on Orgnal Method Identfyng mages through egenspace projecton takes three basc steps. Frst the egenspace must be created usng tranng mages. Net, the tranng mages are projected nto the egenspace. Fnally, the test mages are dentfed by projectng them nto the egenspace and comparng them to the projected tranng mages.. Create Egenspace he followng steps create an egenspace.. Center data: Each of the tranng mages must be centered. Subtractng the mean mage from each of the tranng mages centers the tranng mages as shown n equaton (). he mean mage s a column vector such that each entry s the mean of all correspondng pels of the tranng mages.. Create data matr: Once the tranng mages are centered, they are combned nto a data matr of sze NP, where P s the number of tranng mages and each column s a sngle mage as shown n equaton (3). 3. Create covarance matr: he data matr s multpled by ts transpose to create a covarance matr as shown n equaton (4). 6

13 4. Compute the egenvalues and egenvectors: he egenvalues and correspondng egenvectors are computed for the covarance matr. Ω V ΛV (5) here V s the set of egenvectors assocated wth the egenvalues Λ. 5. Order egenvectors: Order the egenvectors v V accordng to ther correspondng egenvalues λ Λ from hgh to low. Keep only the egenvectors assocated wth nonzero egenvalues. hs matr of egenvectors s the egenspace V, where each column of V s an egenvector.. Project tranng mages [ v v... ] V (6) Each of the centered tranng mages ( ) s projected nto the egenspace. o project an mage nto the egenspace, calculate the dot product of the mage wth each of the ordered egenvectors. v P ~ V (7) herefore, the dot product of the mage and the frst egenvector wll be the frst value n the new vector. he new vector of the projected mage wll contan as many values as egenvectors. 3. Identfy test mages Each test mage s frst mean centered by subtractng the mean mage, and s then projected nto the same egenspace defned by V. y p y m, where m P (8) and 7

14 ~ y V y (9) he projected test mage s compared to every projected tranng mage and the tranng mage that s found to be closest to the test mage s used to dentfy the tranng mage. he mages can be compared usng any number of smlarty measures; the most common s the L norm. I wll dscuss the dfferent smlarty measures n secton 4.3. he followng s an eample of dentfyng mages through egenspace projecton. Let the four mages n Fgure be tranng mages and let the addtonal mage n Fgure be a test mage. he four tranng mages and the mean mage are: m he centered mages are:

15 Fgure. Four tranng mages and one test mage. Combne all the centered tranng mages nto one data matr: Χ Calculate the covarance matr: Ω ΧΧ he ordered nonzero egenvectors of the covarance matr and the correspondng egenvalues are: 9

16 v v v λ 535 λ 5696 λ 3 78 he egenspace s defned by the projecton matr V he four centered tranng mages projected nto egenspace are: ~ V ~ V ~ 3 3 V ~ 4 4 V he test mage vewed as a vector and the centered test mage are:

17 y y he projected test mage s: ~ y V y he L norms are 96, 8, 58 and 449 of the test mage, 3 and y and the tranng mages 4 respectvely. y comparng the L norms, the second tranng mage s found to be closest to the test mage y, therefore the test mage belongng to the same class of mages as the second tranng mage orgnal mages, one sees mage y s most lke. y s dentfed as,. y vewng the. utoral for Snapshot Method of Egenspace Projecton he method outlned above can lead to etremely large covarance matrces. For eample, mages of sze 6464 combne to create a data matr of sze 496P and a covarance matr of sze hs s a problem because calculatng the covarance matr and the egenvectors/egenvalues of the covarance s computatonally demandng. It s known that for a NM matr the mamum number of nonzero egenvectors the matr can have s mn(n,m) [6,7,]. Snce the number of tranng

18 mages ( P ) s usually less than the number of pels ( N ), the most egenvectors/egenvalues that can be found are P. common theorem n lnear algebra states that the egenvalues of X X and X X are the same. Furthermore, the egenvectors of X X are the same as the egenvectors of X X multpled by the matr X and normalzed [6,7,]. Usng ths theorem, the Snapshot method can be used to create the egenspace from a PP matr rather than a NN covarance matr. he followng steps should be followed.. Center data: (Same as orgnal method). Create data matr: (Same as orgnal method) 3. Create covarance matr: he data matr s transpose s multpled by the data matr to create a covarance matr. Ω X X () 4. Compute the egenvalues and egenvectors of O : he egenvalues and correspondng egenvectors are computed for Ω. 5. Compute the egenvectors of egenvectors. Dvde the egenvectors by ther norm. Ω V Λ V () X X V ˆ XV () : Multply the data matr by the v vˆ vˆ (3) 6. Order egenvectors: (Same as orgnal method)

19 he followng s the same eample as used prevously, but the egenspace s calculated usng the Snapshot method. he same tranng and test mages wll be used as shown n Fgure. he revsed covarance matr s: Ω Χ Χ he ordered egenvectors and correspondng nonzero egenvalues of the revsed covarance matr are: v v v λ 535 λ 5696 λ 3 78 he data matr multpled by the egenvectors are: v ˆ v ˆ vˆ elow are the normalzed egenvectors. Note that they are the same egenvectors that were calculated usng the orgnal method. 3

20 v v v Varatons Centerng the mages by subtractng the mean mage s one common method of modfyng the orgnal mages. nother varant s to subtract the mean of each mage from all of the pel values for that mage []. hs varaton smplfes the correlaton calculaton, snce the mages are already mean subtracted. Yet another varaton s to normalze each mage by dvdng each pel value by the norm of the mage, so that the vector has a length of one []. hs varaton smplfes the covarance calculaton to a dot product. n mage cannot be both centered and normalzed, snce these actons counteract the one another. ut an mage can be centered and mean subtracted or mean subtracted and normalzed. For all my work, I use only centered mages. 4

21 3. Fsher Dscrmnants Fsher dscrmnants group mages of the same class and separates mages of dfferent classes. Images are projected from Ndmensonal space (where N s the number of pels n the mage) to C dmensonal space (where C s the number of classes of mages). For eample, consder two sets of ponts n dmensonal space that are projected onto a sngle lne (Fgure a). Dependng on the drecton of the lne, the ponts can ether be med together (Fgure b) or separated (Fgure c). Fsher dscrmnants fnd the lne that best separates the ponts. o dentfy a test mage, the projected test mage s compared to each projected tranng mage, and the test mage s dentfed as the closest tranng mage. 3. Fsher Dscrmnants utoral (Orgnal Method) s wth egenspace projecton, tranng mages are projected nto a subspace. he test mages are projected nto the same subspace and dentfed usng a smlarty measure. What dffers s how the subspace s calculated. Followng are the steps to follow to fnd the Fsher dscrmnants for a set of mages.. Calculate the wthn class scatter matr: he wthn class scatter matr measures the amount of scatter between tems n the same class. For the th class, a scatter matr ( S ) s calculated as the sum of the covarance matrces of the centered mages n that class. 5

22 Fgure. (a) Ponts n dmensonal space. (b) Ponts med when projected onto a lne. (c) Ponts separated when projected onto a lne. where S ( m )( m ) (4) X m s the mean of the mages n the class. he wthn class scatter matr ( S W ) s the sum of all the scatter matrces. C S W S (5) where C s the number of classes.. Calculate the between class scatter matr: he between class scatter matr ( S ) measures the amount of scatter between classes. It s calculated as the sum of the covarance matrces of the dfference between the total mean and the mean of each class. C S n ( m m)( m m) (6) where n s the number of mages n the class, class and m s the mean of all the mages. m s the mean of the mages n the 6

23 3. Solve the generalzed egenvalue problem: Solve for the generalzed egenvectors (V ) and egenvalues ( Λ ) of the wthn class and between class scatter matrces. S V ΛS V (7) W 4. Keep frst Cl egenvectors: Sort the egenvectors by ther assocated egenvalues from hgh to low and keep the frst C egenvectors. hese egenvectors form the Fsher bass vectors. 5. Project mages onto Fsher bass vectors: Project all the orgnal (.e. not centered) mages onto the Fsher bass vectors by calculatng the dot product of the mage wth each of the Fsher bass vectors. he orgnal mages are projected onto ths lne because these are the ponts that the lne has been created to dscrmnate, not the centered mages. Followng s an eample of calculatng the Fsher dscrmnants for a set of mages. Let the twelve mages n Fgure 3 be tranng mages. here are two classes; mages 6 are n the frst class and mages 7 are n the second class. he tranng mages vewed as vectors are:

24 8 Fgure 3. welve tranng mages he scatter matrces are: S

25 S he wthn class scatter matr s: S S S W he mean of each class and the total mean are: m m m

26 he between class scatter matr s: S Snce there are two classes, only one egenvector s kept. he nonzero egenvector and correspondng egenvalue of S V λs V are: W v λ 9.45 he values of the mages projected onto the frst egenvector are shown n able. Fgure 4 shows a plot of the ponts; clearly llustratng the separaton between the two classes. 3. Fsher Dscrmnants utoral (Orthonormal ass Method) wo problems arse when usng Fsher dscrmnants. Frst, the matrces needed for computaton are very large, causng slow computaton tme and possble problems wth

27 able. he values of the mages projected onto the frst egenvector Class Class Fgure 4. Plot of the mages projected onto Fsher bass vectors. numerc precson. Second, snce there are fewer tranng mages than pels, the data matr s rank defcent. It s possble to solve the egenvectors and egenvalues of a rank defcent matr by usng a generalze sngular value decomposton routne, but a smpler soluton ests. smpler soluton s to project the data matr of tranng mages nto an orthonormal bass of sze PP (where P s the number of tranng mages). hs projecton produces a data matr of full rank that s much smaller and therefore decreases computaton tme. he projecton also preserves nformaton so the fnal outcome of Fsher dscrmnants s not affected. Followng are the steps to follow to fnd the Fsher dscrmnants of a set of mages by frst projectng the mages nto any orthonormal bass.. Compute means: Compute the mean of the mages n each class( m )and the total mean of all mages ( m ).. Center the mages n each class: Subtract the mean of each class from the mages n that class. X, X X, ˆ m (8)

28 3. Center the class means: Subtract the total mean from the class means. mˆ m m (9) 4. Create a data matr: Combne the all mages, sdebysde, nto one data matr. 5. Fnd an orthonormal bass for ths data matr: hs can be accomplshed by usng a QR Orthogonaltrangular decomposton or by calculatng the full set of egenvectors of the covarance matr of the tranng data. Let the orthonormal bass be U. 6. Project all centered mages nto the orthonormal bass: Create vectors that are the dot product of the mage and the vectors n the orthonormal bass. ~ U ˆ () 7. Project the centered means nto the orthonormal bass: m ~ U mˆ () 8. Calculate the wthn class scatter matr: he wthn class scatter matr measures the amount of scatter between tems wthn the same class. For the th class a scatter matr ( S ) s calculated as the sum of the covarance matrces of the projected centered mages for that class. S ~ ~ () X he wthn class scatter matr ( S W ) s the sum of all the scatter matrces. C S W S (3) where C s the number of classes.

29 9. Calculate the between class scatter matr: he between class scatter matr ( S ) measures the amount scatter between classes. It s calculated as the sum of the covarance matrces of the projected centered means of the classes, weghted by the number of mages n each class. S C n m ~ m ~ (4) where n s the number of mages n the class.. Solve the generalzed egenvalue problem: Solve for the generalzed egenvectors (V ) and egenvalues ( Λ ) of the wthn class and between class scatter matrces. S V λs V (5) W. Keep the frst Cl egenvectors: Sort the egenvectors by ther assocated egenvalues from hgh to low and keep the frst C egenvectors. hese are the Fsher bass vectors.. Project mages onto egenvectors: Project all the rotated orgnal (.e. Not centered) mages onto the Fsher bass vectors. Frst project the orgnal mages nto the orthonormal bass, and then project these projected mages onto the Fsher bass vectors. he orgnal rotated mages are projected onto ths lne because these are the ponts that the lne has been created to dscrmnate, not the centered mages. he same eample as before wll be calculated usng the orthonormal bass. Let the twelve mages n Fgure 3 be tranng mages. he tranng mages vewed as vectors, the means of each class and the total mean are the same as n the prevous eample. 3

30 4 he centered mages are: ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ he centered class means are: ˆ ˆ m m

31 5 he orthonormal bass calculated by egenspace projecton s: U he centered mages projected nto the orthonormal bass are: ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ 9 8 7

32 6 he centered means projected nto the orthonormal bass are: ~ ~ m m he wthn class scatter matr s: S W Notce that the wthn class scatter matr s a dagonal matr and the values along the dagonal are the egenvalues assocated wth the egenvectors used to create the orthonormal bass. hs occurs because the mages are projected nto ths orthonormal bass before calculatng the wthn class scatter matr. herefore each projected mage s orthogonal to all other projected mages. he between class scatter matr s:

33 S Snce there are two classes, only one egenvector s kept. he nonzero egenvector and correspondng egenvalue of S V λs V are: W v λ he values of the rotated mages projected onto the frst egenvector are shown n able. Fgure 5 shows a plot of the ponts; you can clearly see the separaton between the two classes. 7

34 able. he values of the mages projected onto the frst egenvector Class Class Fgure 5. Plot of the mages projected onto Fsher bass vectors. 8

35 4. Varatons 4. Egenvector Selecton Untl ths pont, when creatng a subspace usng egenspace projecton we use all egenvectors assocated wth nonzero egenvalues. he computaton tme of egenspace projecton s drectly proportonal to the number of egenvectors used to create the egenspace. herefore by removng some porton of the egenvectors computaton tme s decrease. Furthermore, by removng addtonal egenvectors that do not contrbute to the classfcaton of the mage, performance can be mproved. Many varatons of egenvector selecton have been consdered; I wll dscuss fve. hese may be appled ether alone or as part of Fsher dscrmnants.. Standard egenspace projecton: ll egenvectors correspondng to nonzero egenvalues are used to create the subspace.. Remove the last 4% of the egenvectors: Snce the egenvectors are sorted by the correspondng descendng egenvalues, ths method removes the egenvectors that fnd the least amount of varance among the mages. Specfcally, 4% of the egenvectors that fnd the least amount of varance are removed []. 3. Energy dmenson: Rather than use a standard cutoff for all subspaces, ths method uses the mnmum number of egenvectors to guarantee that energy (e) s greater than 9

36 a threshold. typcal threshold s.9. he energy of the th egenvector s the rato of the sum of the frst egenvalues over the sum of all the egenvalues [7] e j k j λ λ j j (3) 4. Stretchng dmenson: nother method of selectng egenvectors based on the nformaton provded by the egenvalues s to calculate the stretch (s) of an egenvector. he stretch of the th egenvector s the rato of the th egenvalue ( λ ) over the mamum egenvalue ( λ ) [7]. common threshold for the stretchng dmenson s.. λ s (3) λ 5. Removng the frst egenvector: he prevous three methods assume that the nformaton n the last egenvectors work aganst classfcaton. hs method assumes that nformaton n the frst egenvector works aganst classfcaton. For eample, lghtng causes consderable varaton n otherwse dentcal mages. Hence, ths method removes the frst egenvector []. Fgure 6 shows the values for energy and stretchng on the FERE dataset. 4. Orderng Egenvectors by LkeImage Dfference Ideally, two mages of the same person should project to the same pont n egenspace. ny dfference between the ponts s unwanted varaton. On the other hand, two mages of dfferent subjects should project to ponts that are as wdely separated as possble. o 3

37 Images Correctly Classfed e67.7% e55.4% e73.79% s.964 e4.76% s.3 e8.4% e77.64% s.555 s.738 e85.% e88.5% e9.9% e97.5% e.% e94.56% e99.3% s.933 s.46 s.6 s.35 s.7 s.35 s.3 s. s Number of Egenvectors Fgure 6. Eample of Energy (e) and Stretchng (s) dmenson of a specfc dataset. capture ths ntuton and use t to order egenvectors, we defne a lkemage dfference (ω ) for each of the k egenvectors []. o defne ω, we wll work wth pars of mages of the same people projected nto egenspace. Let X be tranng mages and Y mages of the correspondng people n the test set ordered such that as follows: j X and y j Y are mages of the same person. Defne ω δ k ω where λ j δ j y j (8) 3

38 When a dfference between mages that ought to match s large relatve to the varance for the dmenson λ then ω s large. Conversely, when the dfference between mages that ought to match s small relatve to the varance, ω s small. Snce the goal s to select egenvectors that brng smlar mages close to each other, we rank the egenvectors n order of ascendng ω and remove some number of the last egenvectors. 4.3 Smlarty & Dstance Measures Once mages are projected nto a subspace, there s the task of determnng whch mages are most lke one another. here are two ways n general to determne how alke mages are. One s to measure the dstance between the mages n Ndmensonal space. he second way s to measure how smlar two mages are. When measurng dstance, one wshes to mnmze dstance, so two mages that are alke produce a small dstance. When measurng smlarty, one wshes to mamze smlarty, so that two lke mages produce a hgh smlarty value. here are many possble smlarty and dstance measures; I wll dscuss fve. L norm: he L norm s also known as the cty block norm or the sum norm. It sums up the absolute dfference between pels[6,]. he L norm of an mage and an mage s: N L (, ) (9) he L norm s a dstance measure. Fgure 7 shows the L dstance between two vectors. 3

39 L norm: he L norm s also known as the Eucldean norm or the Eucldean dstance when ts square root s calculated. It sums up the squared dfference between pels [6,,7]. he L norm of an mage and an mage s: N (, ) ( ) L (3) he L norm s a dstance measure. Fgure 7 shows the L dstance between two vectors. Covarance: Covarance s also known as the angle measure. It calculates the angle between two normalzed vectors. akng the dot product of the normalzed vectors performs ths calculaton [,7]. he covarance between mages and s: cov(, ) (3) Covarance s a smlarty measure. y negatng the covarance value, t becomes a dstance measure []. Fgure 7 shows the covarance between two vectors. Mahalanobs dstance: he Mahalanobs dstance calculates the product of the pels and the egenvalue of a specfc dmenson and sums all these products []. he Mahalanobs dstance between an mage and an mage s: N Mah(, ) C (3) 33

40 Fgure 7. L dstance, L dstance and covarance between two vectors Fgure 8. wo mages wth a negatve correlaton and two that correlate well where C (33) λ Mahalanobs dstance s a dstance measure. Correlaton: Correlaton measures the rate of change between the pels of two mages. It produces a value rangng from to, where a value of ndcates the mages are oppostes of each other and a value of ndcates that the mages are dentcal [7]. he correlaton between an mage and an mage s: corr(, ) N ( µ )( σ σ µ ) (34) where µ s the mean of and σ s the standard devaton of. Fgure 8 shows an eample of two mages wth a negatve correlaton and two that correlate well. 34

41 4.4 re Smlarty Measures the Same Insde and Outsde of Egenspace? n egenspace consstng of all egenvectors assocated wth nonzero egenvalues s an orthonormal bass. n orthonormal bass s a set of vectors where the dot product of any two dstnct vectors s zero and the length of every vector s one. Orthonormal bases have the property that any mage that was used to create the orthonormal bass can be projected nto the full orthonormal bass wth no loss of nformaton. hs means that the mage can be projected nto the orthonormal bass and then converted back to the orgnal mage. For eample, let U be an orthonormal bass and let be an mage used to create U. hen U, where s the mage projected nto U. can be recovered by multplyng by U, U. Gven the fact that no nformaton s lost when projectng specfc mages nto an orthonormal bass, do the values of the smlarty measures change? he answer s that t depends on the smlarty measure. he L norm and correlaton produce dfferent values n the two spaces. Mahalanobs dstance s typcally only used n conjuncton wth egenspace. he L norm and covarance do produce the same value n both spaces; I wll prove ths. heorem 4.: he L norm produces the same value on a par of unprojected vectors and on a par of projected vectors. L (, ) L ( U, U ) (35) 35

42 36 Proof: Let U be an orthonormal bass. Let be a vector such that U. Let be a vector such that U. Now the L norm of s defned n equaton (3) and s the same as: ) ( ) ( (36) he L norm of ) ( s defned as: ), ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ), ( L UU UU UU UU U U U U U U U U L N N Hence, the L norm produces the same value on unprojected vectors and on projected vectors. heorem 4.: Covarance produces the same value on a par of unprojected vectors and on a par of projected vectors. ), cov( ), cov( U U (37) Proof: Let U be an orthonormal bass. Let be a vector such that U and U. Let be a vector such that U and U. he covarance of and s defned n equaton (3) and the covarance of and s defned as: ( ) ( ) U U U U U U U U U ), cov( (38) It s known that U, so

43 cov(, ) (39) U U y theorem 4. U and U. So, cov(, ) cov(, ) (4) Hence, covarance produces the same value on unprojected vectors and on projected vectors. I wll llustrate how each measure behaves wth an eample. Consder two vectors, 7 5, 5. Project these two ponts nto the orthonormal bass U U U Fgure 9 shows ponts,, and. L norm: he L norm produces the same value on unprojected vectors and on projected vectors. Eamne the eample. 37

44 Fgure 9. Plot of ponts,, and. he L norm of L ( ) he L norm of L ( ) s: ( ) ( 7) + ( 5 5) + ( 4 ) s: ( ) ( ) + ( (.44)) + ((.4497) (.3) ) hs eample shows that for ths specfc case they are the same, and the above proof covers the general case. 38

45 Covarance: Covarance produces the same value on unprojected vectors and on projected vectors. Eamne the same eample. he covarance between and s: cov(, ) [ ] he covarance between and s: cov(, ) [ ] hs eample shows that for ths specfc case they are the same and the above proof covers the general case. L norm: he L norm does not produce the same value on unprojected vectors and on projected vectors. Intutvely, two ponts that are located dagonal to each other wll 39

46 produce a larger L dstance than the same ponts rotated to be horzontal to each other. hs can be proven by an eample. he L norm of s: L ( ) he L norm of s: L ( ) (.44) (.4497) (.3) s you can see, these two norms are not the same; hence the L norm does not produce the same value on unprojected vectors and on projected vectors. Correlaton: Correlaton does not produce the same value on unprojected vectors and on projected vectors. Intutvely, Correlaton s effected by the mean of the mages. s mages are rotated ther mean changes, therefore ther correlaton changes. hs can be proven by an eample. he correlaton between and s: µ , µ , σ.87,.566 σ 4

47 N ( )( (, ) µ µ corr σ σ ( )( ) ( )( ) ( )( ) * *.566 (.3333) * * * * he correlaton between and s: ) µ.8643, µ.3675, σ.9485, σ.87 * N ( )( (, ) µ µ corr σ σ ( )( ) ( )( ).9485*5.53 ( )( ) * * *.9485* ) ( 3.775) + ( 3.34).9485*5.53 *.5779 hese two correlatons are not the same; hence the correlaton does not produce the same value on an unprojected vector and on a projected vector. Mahalanobs dstance: ypcally the Mahalanobs dstance s calculated on vectors projected nto egenspace, snce t uses the egenvalues to weght the contrbuton along each as. o llustrate Mahalanobs dstance as measured n egenspace, I wll calculate the Mahalanobs dstance between the projected vectors and. C λ

48 C λ C Mah(, ) C N λ (.668) (5.973*8.769 *.88) + (3.8453*.44*.976) + (.4497*.3*.93) 4

49 5. Eperments I performed four eperments, each on one of two datasets. he frst eperment combnes smlarty measures n an attempt to mprove performance. he second eperment tests whether performance mproves when egenvectors are ordered by lkemage dfference n egenspace projecton. he thrd eperment compares many varatons of egenspace projecton and Fsher dscrmnants on the Cat & Dog dataset. he fourth eperment compares several varatons of egenspace projecton and Fsher dscrmnants on the FERE dataset. 5. Datasets he Cat & Dog dataset and the FERE dataset are used for the eperments. he FERE dataset s restructured for some of the eperments. 5.. he Cat & Dog Dataset he Cat & Dog dataset was created by Mark Stevens and s avalable through the Computer Vson group at Colorado State Unversty. he orgnal dataset of, 6464 pel, grayscale mages conssted of 5 cat mages and 5 golden retrever mages. I collected 5 addtonal cat mages and 5 addtonal mages of dfferent types of dogs. Each mage shows the anmal s head and s taken drectly facng the anmal s face. Fgure shows eample mages from the Cat & Dog dataset. hs s a class classfcaton problem, snce each of the test mages s ether a cat or a dog. 43

50 Fgure. Sample mages from the Cat & Dog dataset Fgure. Sample mages from the FERE dataset 5.. he FERE Dataset Jonathan Phllps at the Natonal Insttute of Standards and echnology made the FERE dataset avalable to our department [,,3]. he FERE database contans mages of 96 ndvduals, wth up to 5 dfferent mages captured for each ndvdual. he mages are separated nto two sets: gallery mages and probes mages. Gallery mages are mages wth known labels, whle probe mages are matched to gallery mages for dentfcaton. he database s broken nto four categores: F: wo mages were taken of an ndvdual, one after the other. One mage s of the ndvdual wth a neutral facal epresson, whle the other s of the ndvdual wth a dfferent epresson. One of the mages s placed nto the gallery fle whle the other s used as a probe. In ths category, the gallery contans 96 mages, and the probe set has 95 mages. 44

51 Duplcate I: he only restrcton of ths category s that the gallery and probe mages are dfferent. he mages could have been taken on the same day or ½ years apart. In ths category, the gallery conssts of the same 96 mages as the F gallery whle the probe set contans 7 mages. fc: Images n the probe set are taken wth a dfferent camera and under dfferent lghtng than the mages n the gallery set. he gallery contans the same 96 mages as the F & Duplcate I galleres, whle the probe set contans 94 mages. Duplcate II: Images n the probe set were taken at least year after the mages n the gallery. he gallery contans 864 mages, whle the probe set has 34 mages. Fgure shows eample mages from the FERE dataset he Restructured FERE Dataset I restructured a porton of the FERE dataset so that there are four mages for each of 6 ndvduals. wo of the pctures are taken on the same day, where one pcture s of the ndvdual wth a neutral facal epresson and the other s wth a dfferent epresson. he other two pctures are taken on a dfferent day wth the same characterstcs. he purpose of ths restructurng s to create a dataset wth more than one tranng mage of each ndvdual to allow testng of Fsher dscrmnants. 5. aggng and Combnng Smlarty Measures Dfferent smlarty measures have been dscussed, but up untl ths pont they have been eamned separately. I wll now eamne combnng some of the smlarty measures 45

52 together n the hopes of mprovng performance. he followng four smlarty measures are eamned: L norm (9), the L norm (3), covarance (3), and the Mahalanobs dstance (3). I test both smple combnatons of the dstance measures and baggng the results of two or more measures usng a votng scheme [,3,9]. 5.. ddng Dstance Measures smple way to combne dstance measures s to add them. In other words, the dstance between two mages s defned as the sum S of the dstances accordng to two or more tradtonal measures: S ( a,..., ah ) a a h (4) Usng S, all combnatons of base metrcs ( L, L, covarance, Mahalonobs) are used to select the nearest gallery mage to each probe mage. he percentage of mages correctly recognzed usng each combnaton on the Duplcate I probe set, s shown n able 5, along wth the recognton rates for the base measures themselves. Of the four base measures, there appears to be a sgnfcant mprovement wth the Mahalanobs dstance. On the surface, 4% seems much better than 33%, 34% or 35%. he best performance of any combned measure s 43% for the S( L, Mahalanobs) and S( L,covarance, Mahalanobs) combnatons. Whle hgher, the dfference does not appear sgnfcant. I used McNemar s test, whch smplfes to the sgn test [,4], to calculate the sgnfcance of dfferences n these results. he McNemar s test calculates how often one algorthm succeeds whle the other algorthm fals. I formulated the followng hypotheses to test sgnfcant dfference n the prevous results. 46

53 able 3. Results of McNemar s test among base measures. lgorthms lgorthm Success/Success Success/Falure Falure/Success P< L L L ngle L Mahalanobs L ngle L Mahalanobs ngle Mahalanobs 5 8. able 4. Results of McNemar s test of the S( L, Mahalanobs) and S( L,covarance, Mahalanobs) combnatons compared to the Mahalanobs dstance lgorthms lgorthm Success/Success Success/Falure Falure/Success P< Mahalanobs S(L +Mahalanobs) Mahalanobs S (L, ng, Mah) Of the four base measures, Mahalanobs dstance outperforms all others, 4% versus 33%, 34% or 35%.. he performance of any combned measures s not statstcally better than the performance of the base measures. 43% for the S( L, Mahalanobs) and S( L,covarance, Mahalanobs) combnatons versus 4% for Mahalanobs dstance. able 3 shows the results of McNemar s test performed for each par of base measures. Note that Mahalanobs dstance always fals less often than the other smlarty measures, ndcated by P <.. Yet, no other measure s found to be dfferent. able 4 shows the results of McNemar s test of the S( L, Mahalanobs) and S( L,covarance, Mahalanobs) combnatons compared to the Mahalanobs dstance. Here no sgnfcant dfference s found between these algorthms, ndcatng that when they do dffer on a partcular mage each s equally lkely to dentfy the mage correctly. 47

54 able 5. Results of addng smlarty measures Classfer Duplcate I L.35 L.33 Covarance.34 Mahalanobs.4 S (L, L ).35 S (L, Covarance).39 S (L, Mahalanobs).43 able 6. Results of baggng smlarty measures Classfer Dup I F L L.33.7 Cov.34.7 Mah.4.74 aggng aggng (best 5) aggng (Weghted) S (L, Covarance).33 S (L, Mahalanobs).4 S (ngle, Covarance).4 S (L, L, Covarance).35 S (L, L, Mahalanobs).4 S (L, Cov, Mah).43 S (L, Cov, Mah).4 S (L, L,Cov, Mah).4 Interestngly, the performance of the combned measures s never less than the performance of ther components evaluated separately. For eample, the performance of S( L, L ) s 35%; ths s better than the performance of L (33%) and the same as L (35%). hese results suggest that L and L are dentfyng the same mages correctly; hence combnng measures does not dentfy any addtonal mages correctly. 5.. Dstance Measure ggregaton he eperment above tested only a smple summaton of dstance measures; one can magne many weghtng schemes for combnng dstance measures that mght outperform smple summaton. Rather than search the space of possble dstance measure combnatons, however, I took a cue from recent work n machne learnng that suggests the best way to combne multple estmators s to apply each estmator ndependently and combne the results by votng [,3,9]. 48

55 For face recognton, ths mples that each dstance measure s allowed to vote for the mage that t beleves s the closest match for a probe. he mage wth the most votes s chosen as the matchng gallery mage. Votng s performed three dfferent ways. aggng: Each classfer s gven one vote as eplaned above. aggng, best of 5: Each classfer votes for the fve gallery mages that most closely match the probe mage. aggng, weghted: Classfers cast fve votes for the closest gallery mage, four votes for the second closest gallery mage, and so on, castng just one vote for the ffth closest mage. able 6 shows the performance of votng for the Duplcate I and F probe sets. On the Duplcate I data, Mahalanobs dstance alone does better than any of the bagged classfers: 4% versus 37% and 38%. On the smpler F probe set, the best performance for a separate classfer s 77% (for L ), and the best performance for the bagged classfers s 78%. he McNemar s test confrms that ths s not a sgnfcant mprovement. In the net secton, I eplore one possble eplanaton for ths lack of mprovement when usng baggng Correlatng Dstance Metrcs s descrbed n [], the falure of votng to mprove performance suggests that the four dstance measures share the same bas. o test ths theory, I correlate the dstances calculated by the four measures over the Duplcate I probe set. Snce each measure s 49

56 defned over a dfferent range, Spearman rank correlaton s used [4]. For each probe mage, the gallery mages are ranked by ncreasng dstance to the probe. hs s done for each par of dstance measures. he result s two rank vectors, one for each dstance measure. Spearman s Rank Correlaton s the correlaton coeffcent for these two vectors. able 7 shows the average correlaton scores. L, covarance and Mahalanobs all correlate very closely to each other, although L correlates less well to covarance and Mahalanobs. hs suggests that there mght be some advantage to combnng L wth covarance or Mahalanobs, but that no combnaton of L, covarance or Mahalanobs s very promsng. hs s consstent wth the scores n able 5, whch show that the combnatons S( L, covarance) and S( L, Mahalanobs) outperform these classfers ndvdually. I also constructed a lst of mages n the F probe set that were grossly msclassfed, n the sense that the matchng gallery mage s not one of the ten closest mages accordng to one or more dstance measures. total of 79 mages are poorly dentfed by at least one dstance measure. able 8 shows the number of mages that are poorly dentfed by all four dstance measures, three dstance measures, two dstance measures, and just one dstance measure. hs table shows that there s shared bas among the classfers, n that they seem to make gross mstakes on the same mages. On the other hand, the errors do not overlap 5

57 able 7. Correlaton between smlarty measures. L L cov Mah L L cov Mah able 8. Number of mages smlarly dentfed poorly Images commonly poorly dentfed # of mages out of 79 % mages by 4 classfers by 3 classfers by classfers by classfer completely, suggestng that some mprovement mght stll be acheved by some combnaton of these dstance measures. 5.3 LkeImage Dfference on the FERE dataset In order to test the performance of lkemage orderng of egenvectors compared to orderng by correspondng egenvalue, I performed two eperments. he frst eperment s on the orgnal FERE dataset and compares results to those of Moon & Phllps. he second eperment s on the restructured dataset, where trals are run and tranng/test data s clearly separated. For each of the 95 probe/gallery matches n the F probe set of the orgnal FERE dataset, I calculate the dfference between the probe and gallery mage n egenspace. hese dfferences are summed together and then dvded by the egenvalue to calculate the lkemage dfference. he smaller ths number s, the better the egenvector should be at matchng mages. he top N egenvalues are selected accordng to the lkemage dfference measure, and the F probe set s reevaluated usng the L norm. Fgure shows the performance scores of the reordered egenvalues compared to the performance of the egenvalues ordered by egenvalue, as performed by Moon & Phllps. able 9 shows the number of mages correctly dentfed by each orderng method and the results 5

Recognizing Faces. Outline

Recognizing Faces. Outline Recognzng Faces Drk Colbry Outlne Introducton and Motvaton Defnng a feature vector Prncpal Component Analyss Lnear Dscrmnate Analyss !"" #$""% http://www.nfotech.oulu.f/annual/2004 + &'()*) '+)* 2 ! &

More information

Feature Reduction and Selection

Feature Reduction and Selection Feature Reducton and Selecton Dr. Shuang LIANG School of Software Engneerng TongJ Unversty Fall, 2012 Today s Topcs Introducton Problems of Dmensonalty Feature Reducton Statstc methods Prncpal Components

More information

Face Recognition University at Buffalo CSE666 Lecture Slides Resources:

Face Recognition University at Buffalo CSE666 Lecture Slides Resources: Face Recognton Unversty at Buffalo CSE666 Lecture Sldes Resources: http://www.face-rec.org/algorthms/ Overvew of face recognton algorthms Correlaton - Pxel based correspondence between two face mages Structural

More information

MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION

MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION Paulo Quntlano 1 & Antono Santa-Rosa 1 Federal Polce Department, Brasla, Brazl. E-mals: quntlano.pqs@dpf.gov.br and

More information

Classifier Selection Based on Data Complexity Measures *

Classifier Selection Based on Data Complexity Measures * Classfer Selecton Based on Data Complexty Measures * Edth Hernández-Reyes, J.A. Carrasco-Ochoa, and J.Fco. Martínez-Trndad Natonal Insttute for Astrophyscs, Optcs and Electroncs, Lus Enrque Erro No.1 Sta.

More information

Machine Learning: Algorithms and Applications

Machine Learning: Algorithms and Applications 14/05/1 Machne Learnng: Algorthms and Applcatons Florano Zn Free Unversty of Bozen-Bolzano Faculty of Computer Scence Academc Year 011-01 Lecture 10: 14 May 01 Unsupervsed Learnng cont Sldes courtesy of

More information

Support Vector Machines

Support Vector Machines /9/207 MIST.6060 Busness Intellgence and Data Mnng What are Support Vector Machnes? Support Vector Machnes Support Vector Machnes (SVMs) are supervsed learnng technques that analyze data and recognze patterns.

More information

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers IOSR Journal of Electroncs and Communcaton Engneerng (IOSR-JECE) e-issn: 78-834,p- ISSN: 78-8735.Volume 9, Issue, Ver. IV (Mar - Apr. 04), PP 0-07 Content Based Image Retreval Usng -D Dscrete Wavelet wth

More information

Optimizing Document Scoring for Query Retrieval

Optimizing Document Scoring for Query Retrieval Optmzng Document Scorng for Query Retreval Brent Ellwen baellwe@cs.stanford.edu Abstract The goal of ths project was to automate the process of tunng a document query engne. Specfcally, I used machne learnng

More information

Modular PCA Face Recognition Based on Weighted Average

Modular PCA Face Recognition Based on Weighted Average odern Appled Scence odular PCA Face Recognton Based on Weghted Average Chengmao Han (Correspondng author) Department of athematcs, Lny Normal Unversty Lny 76005, Chna E-mal: hanchengmao@163.com Abstract

More information

Classifying Acoustic Transient Signals Using Artificial Intelligence

Classifying Acoustic Transient Signals Using Artificial Intelligence Classfyng Acoustc Transent Sgnals Usng Artfcal Intellgence Steve Sutton, Unversty of North Carolna At Wlmngton (suttons@charter.net) Greg Huff, Unversty of North Carolna At Wlmngton (jgh7476@uncwl.edu)

More information

Collaboratively Regularized Nearest Points for Set Based Recognition

Collaboratively Regularized Nearest Points for Set Based Recognition Academc Center for Computng and Meda Studes, Kyoto Unversty Collaboratvely Regularzed Nearest Ponts for Set Based Recognton Yang Wu, Mchhko Mnoh, Masayuk Mukunok Kyoto Unversty 9/1/013 BMVC 013 @ Brstol,

More information

Announcements. Supervised Learning

Announcements. Supervised Learning Announcements See Chapter 5 of Duda, Hart, and Stork. Tutoral by Burge lnked to on web page. Supervsed Learnng Classfcaton wth labeled eamples. Images vectors n hgh-d space. Supervsed Learnng Labeled eamples

More information

UB at GeoCLEF Department of Geography Abstract

UB at GeoCLEF Department of Geography   Abstract UB at GeoCLEF 2006 Mguel E. Ruz (1), Stuart Shapro (2), June Abbas (1), Slva B. Southwck (1) and Davd Mark (3) State Unversty of New York at Buffalo (1) Department of Lbrary and Informaton Studes (2) Department

More information

Appearance-based Statistical Methods for Face Recognition

Appearance-based Statistical Methods for Face Recognition 47th Internatonal Symposum ELMAR-2005, 08-10 June 2005, Zadar, Croata Appearance-based Statstcal Methods for Face Recognton Kresmr Delac 1, Mslav Grgc 2, Panos Latss 3 1 Croatan elecom, Savsa 32, Zagreb,

More information

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points;

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points; Subspace clusterng Clusterng Fundamental to all clusterng technques s the choce of dstance measure between data ponts; D q ( ) ( ) 2 x x = x x, j k = 1 k jk Squared Eucldean dstance Assumpton: All features

More information

Assignment # 2. Farrukh Jabeen Algorithms 510 Assignment #2 Due Date: June 15, 2009.

Assignment # 2. Farrukh Jabeen Algorithms 510 Assignment #2 Due Date: June 15, 2009. Farrukh Jabeen Algorthms 51 Assgnment #2 Due Date: June 15, 29. Assgnment # 2 Chapter 3 Dscrete Fourer Transforms Implement the FFT for the DFT. Descrbed n sectons 3.1 and 3.2. Delverables: 1. Concse descrpton

More information

Mathematics 256 a course in differential equations for engineering students

Mathematics 256 a course in differential equations for engineering students Mathematcs 56 a course n dfferental equatons for engneerng students Chapter 5. More effcent methods of numercal soluton Euler s method s qute neffcent. Because the error s essentally proportonal to the

More information

A Binarization Algorithm specialized on Document Images and Photos

A Binarization Algorithm specialized on Document Images and Photos A Bnarzaton Algorthm specalzed on Document mages and Photos Ergna Kavalleratou Dept. of nformaton and Communcaton Systems Engneerng Unversty of the Aegean kavalleratou@aegean.gr Abstract n ths paper, a

More information

6.854 Advanced Algorithms Petar Maymounkov Problem Set 11 (November 23, 2005) With: Benjamin Rossman, Oren Weimann, and Pouya Kheradpour

6.854 Advanced Algorithms Petar Maymounkov Problem Set 11 (November 23, 2005) With: Benjamin Rossman, Oren Weimann, and Pouya Kheradpour 6.854 Advanced Algorthms Petar Maymounkov Problem Set 11 (November 23, 2005) Wth: Benjamn Rossman, Oren Wemann, and Pouya Kheradpour Problem 1. We reduce vertex cover to MAX-SAT wth weghts, such that the

More information

R s s f. m y s. SPH3UW Unit 7.3 Spherical Concave Mirrors Page 1 of 12. Notes

R s s f. m y s. SPH3UW Unit 7.3 Spherical Concave Mirrors Page 1 of 12. Notes SPH3UW Unt 7.3 Sphercal Concave Mrrors Page 1 of 1 Notes Physcs Tool box Concave Mrror If the reflectng surface takes place on the nner surface of the sphercal shape so that the centre of the mrror bulges

More information

Lecture 4: Principal components

Lecture 4: Principal components /3/6 Lecture 4: Prncpal components 3..6 Multvarate lnear regresson MLR s optmal for the estmaton data...but poor for handlng collnear data Covarance matrx s not nvertble (large condton number) Robustness

More information

TN348: Openlab Module - Colocalization

TN348: Openlab Module - Colocalization TN348: Openlab Module - Colocalzaton Topc The Colocalzaton module provdes the faclty to vsualze and quantfy colocalzaton between pars of mages. The Colocalzaton wndow contans a prevew of the two mages

More information

Term Weighting Classification System Using the Chi-square Statistic for the Classification Subtask at NTCIR-6 Patent Retrieval Task

Term Weighting Classification System Using the Chi-square Statistic for the Classification Subtask at NTCIR-6 Patent Retrieval Task Proceedngs of NTCIR-6 Workshop Meetng, May 15-18, 2007, Tokyo, Japan Term Weghtng Classfcaton System Usng the Ch-square Statstc for the Classfcaton Subtask at NTCIR-6 Patent Retreval Task Kotaro Hashmoto

More information

Cluster Analysis of Electrical Behavior

Cluster Analysis of Electrical Behavior Journal of Computer and Communcatons, 205, 3, 88-93 Publshed Onlne May 205 n ScRes. http://www.scrp.org/ournal/cc http://dx.do.org/0.4236/cc.205.350 Cluster Analyss of Electrcal Behavor Ln Lu Ln Lu, School

More information

Parallelism for Nested Loops with Non-uniform and Flow Dependences

Parallelism for Nested Loops with Non-uniform and Flow Dependences Parallelsm for Nested Loops wth Non-unform and Flow Dependences Sam-Jn Jeong Dept. of Informaton & Communcaton Engneerng, Cheonan Unversty, 5, Anseo-dong, Cheonan, Chungnam, 330-80, Korea. seong@cheonan.ac.kr

More information

Lecture #15 Lecture Notes

Lecture #15 Lecture Notes Lecture #15 Lecture Notes The ocean water column s very much a 3-D spatal entt and we need to represent that structure n an economcal way to deal wth t n calculatons. We wll dscuss one way to do so, emprcal

More information

Multi-stable Perception. Necker Cube

Multi-stable Perception. Necker Cube Mult-stable Percepton Necker Cube Spnnng dancer lluson, Nobuuk Kaahara Fttng and Algnment Computer Vson Szelsk 6.1 James Has Acknowledgment: Man sldes from Derek Hoem, Lana Lazebnk, and Grauman&Lebe 2008

More information

Competitive Sparse Representation Classification for Face Recognition

Competitive Sparse Representation Classification for Face Recognition Vol. 6, No. 8, 05 Compettve Sparse Representaton Classfcaton for Face Recognton Yng Lu Chongqng Key Laboratory of Computatonal Intellgence Chongqng Unversty of Posts and elecommuncatons Chongqng, Chna

More information

Problem Set 3 Solutions

Problem Set 3 Solutions Introducton to Algorthms October 4, 2002 Massachusetts Insttute of Technology 6046J/18410J Professors Erk Demane and Shaf Goldwasser Handout 14 Problem Set 3 Solutons (Exercses were not to be turned n,

More information

A Unified Framework for Semantics and Feature Based Relevance Feedback in Image Retrieval Systems

A Unified Framework for Semantics and Feature Based Relevance Feedback in Image Retrieval Systems A Unfed Framework for Semantcs and Feature Based Relevance Feedback n Image Retreval Systems Ye Lu *, Chunhu Hu 2, Xngquan Zhu 3*, HongJang Zhang 2, Qang Yang * School of Computng Scence Smon Fraser Unversty

More information

X- Chart Using ANOM Approach

X- Chart Using ANOM Approach ISSN 1684-8403 Journal of Statstcs Volume 17, 010, pp. 3-3 Abstract X- Chart Usng ANOM Approach Gullapall Chakravarth 1 and Chaluvad Venkateswara Rao Control lmts for ndvdual measurements (X) chart are

More information

3D vector computer graphics

3D vector computer graphics 3D vector computer graphcs Paolo Varagnolo: freelance engneer Padova Aprl 2016 Prvate Practce ----------------------------------- 1. Introducton Vector 3D model representaton n computer graphcs requres

More information

Synthesizer 1.0. User s Guide. A Varying Coefficient Meta. nalytic Tool. Z. Krizan Employing Microsoft Excel 2007

Synthesizer 1.0. User s Guide. A Varying Coefficient Meta. nalytic Tool. Z. Krizan Employing Microsoft Excel 2007 Syntheszer 1.0 A Varyng Coeffcent Meta Meta-Analytc nalytc Tool Employng Mcrosoft Excel 007.38.17.5 User s Gude Z. Krzan 009 Table of Contents 1. Introducton and Acknowledgments 3. Operatonal Functons

More information

Learning the Kernel Parameters in Kernel Minimum Distance Classifier

Learning the Kernel Parameters in Kernel Minimum Distance Classifier Learnng the Kernel Parameters n Kernel Mnmum Dstance Classfer Daoqang Zhang 1,, Songcan Chen and Zh-Hua Zhou 1* 1 Natonal Laboratory for Novel Software Technology Nanjng Unversty, Nanjng 193, Chna Department

More information

PCA Based Gait Segmentation

PCA Based Gait Segmentation Honggu L, Cupng Sh & Xngguo L PCA Based Gat Segmentaton PCA Based Gat Segmentaton Honggu L, Cupng Sh, and Xngguo L 2 Electronc Department, Physcs College, Yangzhou Unversty, 225002 Yangzhou, Chna 2 Department

More information

y and the total sum of

y and the total sum of Lnear regresson Testng for non-lnearty In analytcal chemstry, lnear regresson s commonly used n the constructon of calbraton functons requred for analytcal technques such as gas chromatography, atomc absorpton

More information

Detection of an Object by using Principal Component Analysis

Detection of an Object by using Principal Component Analysis Detecton of an Object by usng Prncpal Component Analyss 1. G. Nagaven, 2. Dr. T. Sreenvasulu Reddy 1. M.Tech, Department of EEE, SVUCE, Trupath, Inda. 2. Assoc. Professor, Department of ECE, SVUCE, Trupath,

More information

12/2/2009. Announcements. Parametric / Non-parametric. Case-Based Reasoning. Nearest-Neighbor on Images. Nearest-Neighbor Classification

12/2/2009. Announcements. Parametric / Non-parametric. Case-Based Reasoning. Nearest-Neighbor on Images. Nearest-Neighbor Classification Introducton to Artfcal Intellgence V22.0472-001 Fall 2009 Lecture 24: Nearest-Neghbors & Support Vector Machnes Rob Fergus Dept of Computer Scence, Courant Insttute, NYU Sldes from Danel Yeung, John DeNero

More information

Machine Learning. Support Vector Machines. (contains material adapted from talks by Constantin F. Aliferis & Ioannis Tsamardinos, and Martin Law)

Machine Learning. Support Vector Machines. (contains material adapted from talks by Constantin F. Aliferis & Ioannis Tsamardinos, and Martin Law) Machne Learnng Support Vector Machnes (contans materal adapted from talks by Constantn F. Alfers & Ioanns Tsamardnos, and Martn Law) Bryan Pardo, Machne Learnng: EECS 349 Fall 2014 Support Vector Machnes

More information

Wishing you all a Total Quality New Year!

Wishing you all a Total Quality New Year! Total Qualty Management and Sx Sgma Post Graduate Program 214-15 Sesson 4 Vnay Kumar Kalakband Assstant Professor Operatons & Systems Area 1 Wshng you all a Total Qualty New Year! Hope you acheve Sx sgma

More information

Unsupervised Learning and Clustering

Unsupervised Learning and Clustering Unsupervsed Learnng and Clusterng Why consder unlabeled samples?. Collectng and labelng large set of samples s costly Gettng recorded speech s free, labelng s tme consumng 2. Classfer could be desgned

More information

Steps for Computing the Dissimilarity, Entropy, Herfindahl-Hirschman and. Accessibility (Gravity with Competition) Indices

Steps for Computing the Dissimilarity, Entropy, Herfindahl-Hirschman and. Accessibility (Gravity with Competition) Indices Steps for Computng the Dssmlarty, Entropy, Herfndahl-Hrschman and Accessblty (Gravty wth Competton) Indces I. Dssmlarty Index Measurement: The followng formula can be used to measure the evenness between

More information

Image Alignment CSC 767

Image Alignment CSC 767 Image Algnment CSC 767 Image algnment Image from http://graphcs.cs.cmu.edu/courses/15-463/2010_fall/ Image algnment: Applcatons Panorama sttchng Image algnment: Applcatons Recognton of object nstances

More information

Compiler Design. Spring Register Allocation. Sample Exercises and Solutions. Prof. Pedro C. Diniz

Compiler Design. Spring Register Allocation. Sample Exercises and Solutions. Prof. Pedro C. Diniz Compler Desgn Sprng 2014 Regster Allocaton Sample Exercses and Solutons Prof. Pedro C. Dnz USC / Informaton Scences Insttute 4676 Admralty Way, Sute 1001 Marna del Rey, Calforna 90292 pedro@s.edu Regster

More information

Two-Dimensional Supervised Discriminant Projection Method For Feature Extraction

Two-Dimensional Supervised Discriminant Projection Method For Feature Extraction Appl. Math. Inf. c. 6 No. pp. 8-85 (0) Appled Mathematcs & Informaton cences An Internatonal Journal @ 0 NP Natural cences Publshng Cor. wo-dmensonal upervsed Dscrmnant Proecton Method For Feature Extracton

More information

Complex Numbers. Now we also saw that if a and b were both positive then ab = a b. For a second let s forget that restriction and do the following.

Complex Numbers. Now we also saw that if a and b were both positive then ab = a b. For a second let s forget that restriction and do the following. Complex Numbers The last topc n ths secton s not really related to most of what we ve done n ths chapter, although t s somewhat related to the radcals secton as we wll see. We also won t need the materal

More information

Programming in Fortran 90 : 2017/2018

Programming in Fortran 90 : 2017/2018 Programmng n Fortran 90 : 2017/2018 Programmng n Fortran 90 : 2017/2018 Exercse 1 : Evaluaton of functon dependng on nput Wrte a program who evaluate the functon f (x,y) for any two user specfed values

More information

FEATURE EXTRACTION. Dr. K.Vijayarekha. Associate Dean School of Electrical and Electronics Engineering SASTRA University, Thanjavur

FEATURE EXTRACTION. Dr. K.Vijayarekha. Associate Dean School of Electrical and Electronics Engineering SASTRA University, Thanjavur FEATURE EXTRACTION Dr. K.Vjayarekha Assocate Dean School of Electrcal and Electroncs Engneerng SASTRA Unversty, Thanjavur613 41 Jont Intatve of IITs and IISc Funded by MHRD Page 1 of 8 Table of Contents

More information

General Regression and Representation Model for Face Recognition

General Regression and Representation Model for Face Recognition 013 IEEE Conference on Computer Vson and Pattern Recognton Workshops General Regresson and Representaton Model for Face Recognton Janjun Qan, Jan Yang School of Computer Scence and Engneerng Nanjng Unversty

More information

A Fast Content-Based Multimedia Retrieval Technique Using Compressed Data

A Fast Content-Based Multimedia Retrieval Technique Using Compressed Data A Fast Content-Based Multmeda Retreval Technque Usng Compressed Data Borko Furht and Pornvt Saksobhavvat NSF Multmeda Laboratory Florda Atlantc Unversty, Boca Raton, Florda 3343 ABSTRACT In ths paper,

More information

Structure from Motion

Structure from Motion Structure from Moton Structure from Moton For now, statc scene and movng camera Equvalentl, rgdl movng scene and statc camera Lmtng case of stereo wth man cameras Lmtng case of multvew camera calbraton

More information

Human Face Recognition Using Generalized. Kernel Fisher Discriminant

Human Face Recognition Using Generalized. Kernel Fisher Discriminant Human Face Recognton Usng Generalzed Kernel Fsher Dscrmnant ng-yu Sun,2 De-Shuang Huang Ln Guo. Insttute of Intellgent Machnes, Chnese Academy of Scences, P.O.ox 30, Hefe, Anhu, Chna. 2. Department of

More information

Online Detection and Classification of Moving Objects Using Progressively Improving Detectors

Online Detection and Classification of Moving Objects Using Progressively Improving Detectors Onlne Detecton and Classfcaton of Movng Objects Usng Progressvely Improvng Detectors Omar Javed Saad Al Mubarak Shah Computer Vson Lab School of Computer Scence Unversty of Central Florda Orlando, FL 32816

More information

Brave New World Pseudocode Reference

Brave New World Pseudocode Reference Brave New World Pseudocode Reference Pseudocode s a way to descrbe how to accomplsh tasks usng basc steps lke those a computer mght perform. In ths week s lab, you'll see how a form of pseudocode can be

More information

Lecture 5: Multilayer Perceptrons

Lecture 5: Multilayer Perceptrons Lecture 5: Multlayer Perceptrons Roger Grosse 1 Introducton So far, we ve only talked about lnear models: lnear regresson and lnear bnary classfers. We noted that there are functons that can t be represented

More information

Pictures at an Exhibition

Pictures at an Exhibition 1 Pctures at an Exhbton Stephane Kwan and Karen Zhu Department of Electrcal Engneerng Stanford Unversty, Stanford, CA 9405 Emal: {skwan1, kyzhu}@stanford.edu Abstract An mage processng algorthm s desgned

More information

An Image Fusion Approach Based on Segmentation Region

An Image Fusion Approach Based on Segmentation Region Rong Wang, L-Qun Gao, Shu Yang, Yu-Hua Cha, and Yan-Chun Lu An Image Fuson Approach Based On Segmentaton Regon An Image Fuson Approach Based on Segmentaton Regon Rong Wang, L-Qun Gao, Shu Yang 3, Yu-Hua

More information

Machine Learning 9. week

Machine Learning 9. week Machne Learnng 9. week Mappng Concept Radal Bass Functons (RBF) RBF Networks 1 Mappng It s probably the best scenaro for the classfcaton of two dataset s to separate them lnearly. As you see n the below

More information

Computer Animation and Visualisation. Lecture 4. Rigging / Skinning

Computer Animation and Visualisation. Lecture 4. Rigging / Skinning Computer Anmaton and Vsualsaton Lecture 4. Rggng / Sknnng Taku Komura Overvew Sknnng / Rggng Background knowledge Lnear Blendng How to decde weghts? Example-based Method Anatomcal models Sknnng Assume

More information

Improvement of Spatial Resolution Using BlockMatching Based Motion Estimation and Frame. Integration

Improvement of Spatial Resolution Using BlockMatching Based Motion Estimation and Frame. Integration Improvement of Spatal Resoluton Usng BlockMatchng Based Moton Estmaton and Frame Integraton Danya Suga and Takayuk Hamamoto Graduate School of Engneerng, Tokyo Unversty of Scence, 6-3-1, Nuku, Katsuska-ku,

More information

LEAST SQUARES. RANSAC. HOUGH TRANSFORM.

LEAST SQUARES. RANSAC. HOUGH TRANSFORM. LEAS SQUARES. RANSAC. HOUGH RANSFORM. he sldes are from several sources through James Has (Brown); Srnvasa Narasmhan (CMU); Slvo Savarese (U. of Mchgan); Bll Freeman and Antono orralba (MI), ncludng ther

More information

Three supervised learning methods on pen digits character recognition dataset

Three supervised learning methods on pen digits character recognition dataset Three supervsed learnng methods on pen dgts character recognton dataset Chrs Flezach Department of Computer Scence and Engneerng Unversty of Calforna, San Dego San Dego, CA 92093 cflezac@cs.ucsd.edu Satoru

More information

INF Repetition Anne Solberg INF

INF Repetition Anne Solberg INF INF 43 7..7 Repetton Anne Solberg anne@f.uo.no INF 43 Classfers covered Gaussan classfer k =I k = k arbtrary Knn-classfer Support Vector Machnes Recommendaton: lnear or Radal Bass Functon kernels INF 43

More information

Parallel matrix-vector multiplication

Parallel matrix-vector multiplication Appendx A Parallel matrx-vector multplcaton The reduced transton matrx of the three-dmensonal cage model for gel electrophoress, descrbed n secton 3.2, becomes excessvely large for polymer lengths more

More information

Classification of Face Images Based on Gender using Dimensionality Reduction Techniques and SVM

Classification of Face Images Based on Gender using Dimensionality Reduction Techniques and SVM Classfcaton of Face Images Based on Gender usng Dmensonalty Reducton Technques and SVM Fahm Mannan 260 266 294 School of Computer Scence McGll Unversty Abstract Ths report presents gender classfcaton based

More information

Range images. Range image registration. Examples of sampling patterns. Range images and range surfaces

Range images. Range image registration. Examples of sampling patterns. Range images and range surfaces Range mages For many structured lght scanners, the range data forms a hghly regular pattern known as a range mage. he samplng pattern s determned by the specfc scanner. Range mage regstraton 1 Examples

More information

RECOGNIZING GENDER THROUGH FACIAL IMAGE USING SUPPORT VECTOR MACHINE

RECOGNIZING GENDER THROUGH FACIAL IMAGE USING SUPPORT VECTOR MACHINE Journal of Theoretcal and Appled Informaton Technology 30 th June 06. Vol.88. No.3 005-06 JATIT & LLS. All rghts reserved. ISSN: 99-8645 www.jatt.org E-ISSN: 87-395 RECOGNIZING GENDER THROUGH FACIAL IMAGE

More information

Face Detection with Deep Learning

Face Detection with Deep Learning Face Detecton wth Deep Learnng Yu Shen Yus122@ucsd.edu A13227146 Kuan-We Chen kuc010@ucsd.edu A99045121 Yzhou Hao y3hao@ucsd.edu A98017773 Mn Hsuan Wu mhwu@ucsd.edu A92424998 Abstract The project here

More information

Orthogonal Complement Component Analysis for Positive Samples in SVM Based Relevance Feedback Image Retrieval

Orthogonal Complement Component Analysis for Positive Samples in SVM Based Relevance Feedback Image Retrieval Orthogonal Complement Component Analyss for ostve Samples n SVM Based Relevance Feedback Image Retreval Dacheng Tao and Xaoou Tang Department of Informaton Engneerng The Chnese Unversty of Hong Kong {dctao2,

More information

Correlative features for the classification of textural images

Correlative features for the classification of textural images Correlatve features for the classfcaton of textural mages M A Turkova 1 and A V Gadel 1, 1 Samara Natonal Research Unversty, Moskovskoe Shosse 34, Samara, Russa, 443086 Image Processng Systems Insttute

More information

Helsinki University Of Technology, Systems Analysis Laboratory Mat Independent research projects in applied mathematics (3 cr)

Helsinki University Of Technology, Systems Analysis Laboratory Mat Independent research projects in applied mathematics (3 cr) Helsnk Unversty Of Technology, Systems Analyss Laboratory Mat-2.08 Independent research projects n appled mathematcs (3 cr) "! #$&% Antt Laukkanen 506 R ajlaukka@cc.hut.f 2 Introducton...3 2 Multattrbute

More information

CS 534: Computer Vision Model Fitting

CS 534: Computer Vision Model Fitting CS 534: Computer Vson Model Fttng Sprng 004 Ahmed Elgammal Dept of Computer Scence CS 534 Model Fttng - 1 Outlnes Model fttng s mportant Least-squares fttng Maxmum lkelhood estmaton MAP estmaton Robust

More information

User Authentication Based On Behavioral Mouse Dynamics Biometrics

User Authentication Based On Behavioral Mouse Dynamics Biometrics User Authentcaton Based On Behavoral Mouse Dynamcs Bometrcs Chee-Hyung Yoon Danel Donghyun Km Department of Computer Scence Department of Computer Scence Stanford Unversty Stanford Unversty Stanford, CA

More information

Related-Mode Attacks on CTR Encryption Mode

Related-Mode Attacks on CTR Encryption Mode Internatonal Journal of Network Securty, Vol.4, No.3, PP.282 287, May 2007 282 Related-Mode Attacks on CTR Encrypton Mode Dayn Wang, Dongda Ln, and Wenlng Wu (Correspondng author: Dayn Wang) Key Laboratory

More information

Improving Web Image Search using Meta Re-rankers

Improving Web Image Search using Meta Re-rankers VOLUME-1, ISSUE-V (Aug-Sep 2013) IS NOW AVAILABLE AT: www.dcst.com Improvng Web Image Search usng Meta Re-rankers B.Kavtha 1, N. Suata 2 1 Department of Computer Scence and Engneerng, Chtanya Bharath Insttute

More information

High Dimensional Data Clustering

High Dimensional Data Clustering Hgh Dmensonal Data Clusterng Charles Bouveyron 1,2, Stéphane Grard 1, and Cordela Schmd 2 1 LMC-IMAG, BP 53, Unversté Grenoble 1, 38041 Grenoble Cede 9, France charles.bouveyron@mag.fr, stephane.grard@mag.fr

More information

USING GRAPHING SKILLS

USING GRAPHING SKILLS Name: BOLOGY: Date: _ Class: USNG GRAPHNG SKLLS NTRODUCTON: Recorded data can be plotted on a graph. A graph s a pctoral representaton of nformaton recorded n a data table. t s used to show a relatonshp

More information

CS434a/541a: Pattern Recognition Prof. Olga Veksler. Lecture 15

CS434a/541a: Pattern Recognition Prof. Olga Veksler. Lecture 15 CS434a/541a: Pattern Recognton Prof. Olga Veksler Lecture 15 Today New Topc: Unsupervsed Learnng Supervsed vs. unsupervsed learnng Unsupervsed learnng Net Tme: parametrc unsupervsed learnng Today: nonparametrc

More information

Performance Evaluation of Information Retrieval Systems

Performance Evaluation of Information Retrieval Systems Why System Evaluaton? Performance Evaluaton of Informaton Retreval Systems Many sldes n ths secton are adapted from Prof. Joydeep Ghosh (UT ECE) who n turn adapted them from Prof. Dk Lee (Unv. of Scence

More information

Graph-based Clustering

Graph-based Clustering Graphbased Clusterng Transform the data nto a graph representaton ertces are the data ponts to be clustered Edges are eghted based on smlarty beteen data ponts Graph parttonng Þ Each connected component

More information

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision SLAM Summer School 2006 Practcal 2: SLAM usng Monocular Vson Javer Cvera, Unversty of Zaragoza Andrew J. Davson, Imperal College London J.M.M Montel, Unversty of Zaragoza. josemar@unzar.es, jcvera@unzar.es,

More information

S1 Note. Basis functions.

S1 Note. Basis functions. S1 Note. Bass functons. Contents Types of bass functons...1 The Fourer bass...2 B-splne bass...3 Power and type I error rates wth dfferent numbers of bass functons...4 Table S1. Smulaton results of type

More information

Face Recognition Based on SVM and 2DPCA

Face Recognition Based on SVM and 2DPCA Vol. 4, o. 3, September, 2011 Face Recognton Based on SVM and 2DPCA Tha Hoang Le, Len Bu Faculty of Informaton Technology, HCMC Unversty of Scence Faculty of Informaton Scences and Engneerng, Unversty

More information

Integrated Expression-Invariant Face Recognition with Constrained Optical Flow

Integrated Expression-Invariant Face Recognition with Constrained Optical Flow Integrated Expresson-Invarant Face Recognton wth Constraned Optcal Flow Chao-Kue Hseh, Shang-Hong La 2, and Yung-Chang Chen Department of Electrcal Engneerng, Natonal Tsng Hua Unversty, Tawan 2 Department

More information

Optimal Workload-based Weighted Wavelet Synopses

Optimal Workload-based Weighted Wavelet Synopses Optmal Workload-based Weghted Wavelet Synopses Yoss Matas School of Computer Scence Tel Avv Unversty Tel Avv 69978, Israel matas@tau.ac.l Danel Urel School of Computer Scence Tel Avv Unversty Tel Avv 69978,

More information

Skew Angle Estimation and Correction of Hand Written, Textual and Large areas of Non-Textual Document Images: A Novel Approach

Skew Angle Estimation and Correction of Hand Written, Textual and Large areas of Non-Textual Document Images: A Novel Approach Angle Estmaton and Correcton of Hand Wrtten, Textual and Large areas of Non-Textual Document Images: A Novel Approach D.R.Ramesh Babu Pyush M Kumat Mahesh D Dhannawat PES Insttute of Technology Research

More information

Support Vector Machines

Support Vector Machines Support Vector Machnes Decson surface s a hyperplane (lne n 2D) n feature space (smlar to the Perceptron) Arguably, the most mportant recent dscovery n machne learnng In a nutshell: map the data to a predetermned

More information

Robust Face Recognition Using Eigen Faces and Karhunen-Loeve Algorithm

Robust Face Recognition Using Eigen Faces and Karhunen-Loeve Algorithm Robust Face Recognton Usng Egen Faces and Karhunen-Loeve Algorthm Parvnder S. Sandhu, Iqbaldeep Kaur, Amt Verma, Prateek Gupta Abstract The current research paper s an mplementaton of Egen Faces and Karhunen-Loeve

More information

Outline. Self-Organizing Maps (SOM) US Hebbian Learning, Cntd. The learning rule is Hebbian like:

Outline. Self-Organizing Maps (SOM) US Hebbian Learning, Cntd. The learning rule is Hebbian like: Self-Organzng Maps (SOM) Turgay İBRİKÇİ, PhD. Outlne Introducton Structures of SOM SOM Archtecture Neghborhoods SOM Algorthm Examples Summary 1 2 Unsupervsed Hebban Learnng US Hebban Learnng, Cntd 3 A

More information

An efficient method to build panoramic image mosaics

An efficient method to build panoramic image mosaics An effcent method to buld panoramc mage mosacs Pattern Recognton Letters vol. 4 003 Dae-Hyun Km Yong-In Yoon Jong-Soo Cho School of Electrcal Engneerng and Computer Scence Kyungpook Natonal Unv. Abstract

More information

AP PHYSICS B 2008 SCORING GUIDELINES

AP PHYSICS B 2008 SCORING GUIDELINES AP PHYSICS B 2008 SCORING GUIDELINES General Notes About 2008 AP Physcs Scorng Gudelnes 1. The solutons contan the most common method of solvng the free-response questons and the allocaton of ponts for

More information

Edge Detection in Noisy Images Using the Support Vector Machines

Edge Detection in Noisy Images Using the Support Vector Machines Edge Detecton n Nosy Images Usng the Support Vector Machnes Hlaro Gómez-Moreno, Saturnno Maldonado-Bascón, Francsco López-Ferreras Sgnal Theory and Communcatons Department. Unversty of Alcalá Crta. Madrd-Barcelona

More information

Shape Representation Robust to the Sketching Order Using Distance Map and Direction Histogram

Shape Representation Robust to the Sketching Order Using Distance Map and Direction Histogram Shape Representaton Robust to the Sketchng Order Usng Dstance Map and Drecton Hstogram Department of Computer Scence Yonse Unversty Kwon Yun CONTENTS Revew Topc Proposed Method System Overvew Sketch Normalzaton

More information

Biostatistics 615/815

Biostatistics 615/815 The E-M Algorthm Bostatstcs 615/815 Lecture 17 Last Lecture: The Smplex Method General method for optmzaton Makes few assumptons about functon Crawls towards mnmum Some recommendatons Multple startng ponts

More information

Exercises (Part 4) Introduction to R UCLA/CCPR. John Fox, February 2005

Exercises (Part 4) Introduction to R UCLA/CCPR. John Fox, February 2005 Exercses (Part 4) Introducton to R UCLA/CCPR John Fox, February 2005 1. A challengng problem: Iterated weghted least squares (IWLS) s a standard method of fttng generalzed lnear models to data. As descrbed

More information

Some material adapted from Mohamed Younis, UMBC CMSC 611 Spr 2003 course slides Some material adapted from Hennessy & Patterson / 2003 Elsevier

Some material adapted from Mohamed Younis, UMBC CMSC 611 Spr 2003 course slides Some material adapted from Hennessy & Patterson / 2003 Elsevier Some materal adapted from Mohamed Youns, UMBC CMSC 611 Spr 2003 course sldes Some materal adapted from Hennessy & Patterson / 2003 Elsever Scence Performance = 1 Executon tme Speedup = Performance (B)

More information

Reducing Frame Rate for Object Tracking

Reducing Frame Rate for Object Tracking Reducng Frame Rate for Object Trackng Pavel Korshunov 1 and We Tsang Oo 2 1 Natonal Unversty of Sngapore, Sngapore 11977, pavelkor@comp.nus.edu.sg 2 Natonal Unversty of Sngapore, Sngapore 11977, oowt@comp.nus.edu.sg

More information

A Fast Visual Tracking Algorithm Based on Circle Pixels Matching

A Fast Visual Tracking Algorithm Based on Circle Pixels Matching A Fast Vsual Trackng Algorthm Based on Crcle Pxels Matchng Zhqang Hou hou_zhq@sohu.com Chongzhao Han czhan@mal.xjtu.edu.cn Ln Zheng Abstract: A fast vsual trackng algorthm based on crcle pxels matchng

More information

SHAPE RECOGNITION METHOD BASED ON THE k-nearest NEIGHBOR RULE

SHAPE RECOGNITION METHOD BASED ON THE k-nearest NEIGHBOR RULE SHAPE RECOGNITION METHOD BASED ON THE k-nearest NEIGHBOR RULE Dorna Purcaru Faculty of Automaton, Computers and Electroncs Unersty of Craoa 13 Al. I. Cuza Street, Craoa RO-1100 ROMANIA E-mal: dpurcaru@electroncs.uc.ro

More information