LAMP: 3D Layered, Adaptive-resolution and Multiperspective Panorama - a New Scene Representation

Size: px
Start display at page:

Download "LAMP: 3D Layered, Adaptive-resolution and Multiperspective Panorama - a New Scene Representation"

Transcription

1 Submission o Special Issue of CVIU on Model-based and Image-based 3D Scene Represenaion for Ineracive Visualizaion LAMP: 3D Layered, Adapive-resoluion and Muliperspecive Panorama - a New Scene Represenaion Zhigang Zhu Deparmen of Compuer Science The Ciy College, The Ciy Universiy of New York, New York, NY Allen R. Hanson Deparmen of Compuer Science Universiy of Massachuses Amhers, Amhers, MA Shor Running Tile: 3D LAMP Represenaion Conac Informaion: Prof. Zhigang Zhu Deparmen of Compuer Science The Ciy College of New York /CUNY Conven Avenue and 138h Sree, New York, NY Tel: (212) Fax: (212) zhu@cs.ccny.cuny.edu URL: hp://www-cs.engr.ccny.cuny.edu/~zhu/

2 Absrac A compac visual represenaion, called he 3D Layered, Adapive-resoluion and Muliperspecive Panorama (LAMP), is proposed for represening large-scale 3D scenes wih large variaions of dephs and obvious occlusions. Two kinds of 3D LAMP represenaions are proposed: he relief-like LAMP and he image-based LAMP. Boh ypes of LAMPs concisely represen almos all he informaion from a long image sequence. Mehods o consruc LAMP represenaions from video sequences wih dominan ranslaion are provided. The relief-like LAMP is basically a single exended muli-perspecive panoramic view image. Each pixel has a pair of exure and deph values, bu each pixel may also have muliple pairs of exure-deph values o represen occlusion in layers, in addiion o adapive resoluion changing wih deph. The image-based LAMP, on he oher hand, consiss of a se of muli-perspecive layers, each of which has a pair of 2D exure and deph maps, bu wih adapive ime-sampling scales depending on dephs of scene poins. Several examples of 3D LAMP consrucion for real image sequences are given. The 3D LAMP is a concise and powerful represenaion for image-based rendering. Keywords: image-based modeling and rendering, layered represenaion, muli-resoluion, muli-image processing, spaio-emporal image, epipolar plane image 1

3 1. Inroducion Relaed Work D LAMP: Overview of Our Approach Organizaion of he Paper Basic Panoramic Geomery Relief-like 3D LAMP Represenaion Image-Based 3D LAMP Represenaion Deph Layering Temporal Re-Sampling Deph, Occlusion and Resoluion Recovery Deph Recovery: he Panoramic EPI-Based Approach Deph Boundary Localizaion Occlusion Recovery Resoluion Recovery LAMP Consrucion and Rendering: Experimenal Resuls D LAMP Consrucion Resuls Exended Panoramic Image (XPI) Represenaion LAMP-Based Rendering Comparisons and Discussions Layered Represenaion Full Perspecive, Full Orhogonal, Muli-Perspecive and Mulivalued Represenaions Conclusions Acknowledgemens References

4 1. INTRODUCTION The problem of modeling and rendering real 3D scenes has received increasing aenion in recen years in boh he compuer vision and compuer graphics communiies [1-3, 5, 6, 8, 17-19, 22, 23, 27]. In order o build a visual represenaion from image sequences for re-rendering real 3D naural scenes, here are wo challenging issues ha need o be solved: he correspondence problem beween wo (or muliple) views, and a suiable geomeric represenaion of large scale scenes. Usually a suiable visual represenaion will ease he correspondence problem. Many of he successful image-based modeling and rendering approaches [1-3, 5, 6, 8-10] have ried o simplify or avoid he correspondence problem by using 2D image inerpolaion or video regisraion/mosaicing. On he oher hand, more general approaches need sophisicaed vision algorihms, such as in muli-view sereo [17, 18] or general moion analysis [10, 11, 22, 23] of an image sequence. Three classes of image-based represenaions have been proposed o represen video sequences of large-scale 3D scenes: panoramic (mosaic) represenaions, muli-view represenaion and layered represenaions. Building and mainaining a suiable visual represenaion of a large-scale scene remains an imporan research opic in image-based modeling and rendering Relaed Work Mosaic-based approach - A video mosaic is a composie of images creaed by regisering overlapping frames. Many of he curren successful image mosaic algorihms, however, only generae 2D mosaics from a camera roaing around is nodal poin [1,2,3]. Creaing muliperspecive sereo panoramas from one roaing camera off is nodal poin was proposed by Ishiguro, e al [4], Peleg, e al [5], and Shum & Szeliski [6]. A sysem for creaing a global view for visual navigaion by pasing ogeher columns from images aken by a smoohly ranslaing camera (comprising only a verical sli) has been proposed by Zheng & Tsuji [7]. The moving sli paradigm was laer used as he basis of he 2D manifold projecion approach for image mosaicing [8], he muliple-cener-of-projecion approach for image-based rendering [9], he epipolar plane analysis echniques for 3D reconsrucion [15], and he parallel-perspecive view inerpolaion mehods for sereo mosaicing [16, 27]. Muli-perspecive panoramas (or mosaics) exhibi very aracive properies for visual represenaion and epipolar geomery; however, 3D 3

5 recovery from sereo mosaics sill faces he same problems as radiional sereo mehods - he correspondence problem and occlusion handling. The 3D informaion for regions ha are occluded in one of he views is usually difficul o obain since no correspondences can be esablished for hose regions. Layered represenaion - In a layered represenaion, a se of deph surfaces is firs esimaed from an image sequence from a single camera, and hen combined o generae a new view. Wang and Adelson [10] addressed he problem as he compuaion of 2D affine moion models and a se of affine suppor layers from an image sequence. The layered represenaion ha hey proposed consiss of hree maps in each layer: a mosaiced inensiy map, an alpha map, and a velociy map. Occlusion boundaries are represened as disconinuiies in a layer's alpha map (opaciy). This represenaion is a good choice for image compression of a video sequence and for limied image synhesis of seleced layers. Recenly, Ke & Kanade [22] proposed a subspace approach o reliably exrac planar layers from images by aking advanage of he fac ha homographies induced by planar paches in he scene form a low dimensional linear space. Occlusion regions are excluded from heir layered represenaion. Sawhney and Ayer [11] proposed a muliple moion esimaion mehod based on a Minimum Descripion Lengh (MDL) principle. However, heir algorihms are compuaionally expensive and require a combinaorial search o deermine he correc number of layers and he "projecive deph" of each poin in a layer. Occlusion regions are no recovered in heir layered model. Baker, Szeliski & Anandan [12] proposed a framework for exracing srucure from sereo pairs and represened a scene as a collecion of approximaely planar layers. Each layer consiss of an explici 3D plane equaion, a exure map (a sprie), and a map wih deph offses relaive o he plane. The iniial esimaes of he layers are recovered using echniques from parameric moion esimaion and hen refined using a re-synhesis algorihm which akes ino accoun boh occlusion and mixed pixels. More recen work on 3D layer exracion [23] uses an inegraed Bayesian approach o auomaically deermine he number of layers and he assignmen of individual pixels o layers. For more complex geomery, a layered deph image (LDI) has been proposed [13] which is a represenaion of a scene from a single inpu camera view, bu wih muliple pixels along each line of sigh. 4

6 Muli-view approach - Raher han consrucing a single mosaic from a sequence of images, muli-view approaches represen a scene by muliple images wih deph and exure. Chang and Zakhor [17] proposed a mehod o obain deph informaion of some pre-specified reference frames of an image sequence capured by an uncalibraed camera scanning a saionary scene, hen o ransform he poins of reference frames ono an image of he desired virual viewpoin. However, reference frames were chosen quie heurisically; a synhesized image from a viewpoin far away from ha of he reference frames leads o erroneous resuls since occluded or uncovered regions canno be well represened. Chang and Zakhor laer exended his work o a mulivalued represenaion (MVR) [26], which is auomaically consruced wih respec o a single reference frame from a se of dense deph maps. They poined ou ha occlusion levels come naurally from he dense deph informaion and argued ha because of visibiliy limiaions, real-world scenes ypically do no have more han hree occlusion levels. This disinguishes heir MVR from he layered represenaions, which usually have a much larger number of layers when using affine moion o group regions. Szeliski [18] presened a new approach o compuing dense deph and moion esimaes from muliple images. Raher han esimaing a single deph or moion map, a deph or moion map is associaed wih each inpu image (or some subse of hem). Furhermore, a moion compaibiliy consrain is used o ensure consisency beween hese esimaes, and occlusion relaionships are mainained by compuing pixel visibiliy D LAMP: Overview of Our Approach The goal of our work is o consruc a layered and panoramic represenaion of a large-scale 3D scene wih large variaions of dephs and obvious occlusions from primarily ranslaing video sequences. Our approach in his paper is based on a muli-perspecive panoramic view image (PVI) [7] and a se of epipolar plane images (EPIs) [14] exraced from a long image sequence, ogeher wih a panoramic deph map generaed by analyzing he EPIs [15, 21]. To accomplish his, we need o solve four problems: 1) generaing seamless PVIs and EPIs from video under more general moion han pure ranslaion in a large and real scene, 2) analyzing he huge amoun of daa in EPIs robusly and efficienly o obain dense deph informaion, 3) enhancing resoluion and recovering occlusions in a parallel-perspecive PVI represenaion, and 4) represening a large scale 3D scene wih differen dephs and occlusions efficienly and 5

7 compacly. While he firs wo issues are very imporan in consrucing a 3D model of a scene, hey have been discussed in our previous work [15, 21], so his paper will mainly focus on he las wo issues. We propose a new compac represenaion - 3D Layered, Adapive-resoluion and Muliperspecive Panorama (LAMP). The moivaion for layering is o represen occluded regions and he differen spaial resoluions of objecs wih differen deph ranges; meanwhile he model is represened in he form of a seamless muli-perspecive mosaic (panorama) wih viewpoins spanning a large disance. Two kinds of 3D LAMP represenaions are consruced: he relief-like LAMP and he 2D image-based LAMP. The relief-like LAMP is basically a single, exended, muli-perspecive PVI wih boh exure and deph values, bu each pixel has muliple values of exure-deph pairs o represen resuls of occlusion recovery and resoluion enhancemen. The image-based LAMP, on he oher hand, consiss of a se of muli-perspecive layers, each of which has boh exure and deph maps wih adapive ime-sampling scales depending on dephs of scene poins. The LAMP represenaion is relaed o such represenaions as he panoramic view image (PVI) [7, 15], sprie [12], and he layered deph image (LDI) [13]. However, i is more han a muli-perspecive PVI in ha deph, adapive-resoluion and occlusion are added in our represenaion. I is differen from he sprie (or he layered deph image) since he laer is a view of a scene from a single inpu camera view and is wihou adapive image resoluion. The 3D LAMP represenaion is capable of synhesizing images of new views wihin a reasonably resriced bu arbirary moving space, as is inensiy and deph maps conain almos all he informaion ha could be obained from an image sequence Organizaion of he Paper This paper is organized as follows. In secion 2, we describe he basic panoramic represenaion wih parallel-perspecive geomery, including boh exure and deph maps. Based on his supporing panoramic represenaion, we propose a relief-like LAMP represenaion in Secion 3, which feaures muliple layers, adapive resoluion, muli-perspecive, and panoramic views. In Secion 4, an image-based LAMP represenaion is presened wih jus a se of 2D images in order o simplify he 3D scene represenaion. Seps o conver a relief-like LAMP o an imagebased LAMP are provided. Secion 5 discusses a unique EPI-based approach o consruc a relief-like LAMP. This approach does no explicily exrac feaures, rack loci, or mach 6

8 correspondences. Insead, a spaio-emporal-frequency domain cross-analysis is performed o creae dense deph maps, o localize accurae deph boundaries, and o exrac occluded regions ha canno be seen in he seleced panoramic view. In Secion 6, experimenal resuls and some pracical consideraions are given for 3D LAMP consrucion, and 3D rendering based on LAMP represenaions are discussed. In Secion 7, we make wo comparisons. Firs, we compare our LAMP represenaion based on parallel-perspecive projecion wih hose represenaions having full perspecive or full orhogonal projecions. Second, we compare our image-based LAMP represenaion wih several exising layered represenaions. The las secion is a summary of his work. 2. BASIC PANORAMIC GEOMETRY For compleeness, we give a brief inroducion o he consrucions and represenaions of Panoramic View Images (PVIs) and Epipolar Plane Images (EPIs) and o he reconsrucion of a supporing PVI surface - a panoramic map wih boh deph and exure - from PVIs and EPIs. y Z y PVI (x=0) (a) X Y x O f V (b) o x EPI (y=0) Fig. 1. ST image model. A PVI is a y image inside he xy cube and an EPI is an x image Le us consider he siuaion in which a model of a large-scale scene will be consruced from a long and dense video sequence capured by a camera moving in a sraigh line whose line of sigh (i.e., opical axis) is perpendicular o he moion (Fig. 1). The resuling sequence obeys he following spaio-emporal (ST) perspecive projecion model X + V Y x ( ) = f, y( ) = f (1) Z Z where (X,Y,Z) represens he 3D coordinae of a poin a ime =0, f is he camera focal lengh, and V is is speed. A feaure poin (x,y) forms a sraigh locus and is deph value is 7

9 V Vd D = Z = f = f (2) v dx where v = dx / d (3) is he slope of he sraigh locus. In oher words, wo kinds of useful 2D ST images can be exraced: (1) Panoramic View Images (PVIs) - he y- inersecions in he xy cube (Fig. 1), which is a parallel-perspecive image including mos of he 2D informaion of he scene from muliple viewpoins, and (2) Epipolar Plane Images (EPIs) - he x- inersecions in he xy cube, whose ST locus orienaions represen dephs of scene poins. Fig. 1 illusraes he cenral PVI and one of he EPIs. (a) Frame 0 Frame 200 Frame 400 Frame 600 Frame 1000 (b) (c) Fig. 2. A few frames of he sabilized Main Building (MB) sequence in (a), and a pair of sereo PVIs: (b) x = 0 and (c) x= -56. Whie lines indicae maches. In real scenes, he moion of a camera usually will be composed of a dominan ranslaion plus unpredicable variaions. In addiion o deph recovery and occlusion handling, a pre-processing sep is needed o generae a sabilized image sequence wih only a 1D ranslaion by using image sabilizaion and image mosaicing echniques [15, 21]. Fig. 2 shows wo PVIs ha are exraced from x=0 and x=-56 of a 128*128*1024 xy image cube of a building scene - he sabilized Main Building (MB) sequence. The MB sequence was capured by a camcorder mouned on a ripod carried by a moving vehicle. Oher han he small vibraion of he vehicle on a normal oudoor 8

10 road, he ranslaional velociy was mosly consan. Fig. 2a shows a few frames afer video sabilizaion; image ransformaions for removing he camera vibraion produce blank regions (shown in black) around he borders of he sabilized frames, paricularly visible on he righ sides of Frame 0 and Frame The camera vibraion is also demonsraed by he irregular boundaries of he mosaic in Fig. 7. Muli-perspecive PVIs provide a compac represenaion for a large-scale scene, and sereo PVIs can be used o esimae he deph informaion of he scene [7, 6, 16, 27]. In panoramic sereo (Eq. (2), Fig. 2), he "dispariy" dx is fixed; and he disance D is proporional o d, he sereoscopic displacemen in. This indicaes ha deph resoluions are he same for differen dephs. However, we sill face wo problems in order o use sereo PVIs o recover 3D informaion - he correspondence problem and he occlusion problem. EPI (y=9) PVI (x=0) y x o occluded side PVI (x=-56) Fig. 3. Each EPI is a composie image of a scanline from all possible views of an image sequence, and includes informaion abou deph, occlusion and spaial resoluion of 3D scene poins Our soluion o hese wo problems is o effecively use he informaion ha is coninuous beween wo views, i.e., he epipolar plane images, o obain a supporing PVI surface - a panoramic view image wih boh exure and deph maps, and hen he occluded regions ha canno be seen from he PVI. Fig. 3 shows an EPI from which he occluded (and side) regions as well as dephs can be recovered; he algorihms o exrac deph and o recover occlusion/resoluion will be discussed in Secion 5. The supporing PVI surface is a 2D panoramic view image wih boh exure and deph map (Fig. 4a) and is he base for our LAMP represenaion. The exure map is a y- inersecion plane in he xy image cube, and he deph map has deph values for each pixel. In principle, we can exrac he supporing PVI surface from 9

11 any x locaion inside he xy cube. In his paper, we usually selec a cenral PVI (wih x = 0) ha has orhogonal parallel rays along he direcion of moion. For consrucing a dense deph map for he supporing PVI, we need o process all he EPIs inside he xy cube. While he 1D moion model will be used o develop our EPI-based approach in his paper, LAMP represenaions can be consruced under a more general [15, 16, 27] or differen moion [5,6], and/or using oher approaches [10, 11, 12, 22, 23]. (a) Texure and deph maps of he supporing PVI surface (back of he relief). y -x y -x (b) Texure and deph maps of he relief surface (fron of he relief). Fig. 4. The base layer of he relief-like LAMP represenaion for he MB sequence. 3. RELIEF-LIKE 3D LAMP REPRESENTATION Based on he represenaion of a muli-perspecive supporing PVI surface (conaining boh exure map and deph map), we propose a compac and comprehensive represenaion called he 3D LAMP- Layered, Adapive-resoluion and Muli-perspecive Panorama. We will explain he LAMP model by an illusraive example shown in Fig. 5 using a simple 1D scene. Recall ha a 3D ST image (which is a 2D EPI for he 1D scene in Fig. 5) includes everyhing from an image 10

12 sequence. Firs, we will give a basic LAMP represenaion - he relief-like 3D LAMP - ha is direcly carved from he 3D ST image (xy image). The relief-like 3D LAMP basically cus ou some essenial par of he 3D ST image depending on he deph value for each pixel. I has he following four properies ha make i very suiable for image-based modeling and rendering: (a) d 1 s 1 d 2 s 2 d 3 ( o 1 o 2 ) 1D scene Z view direcions for LAMP d 4 X O V x side layer s 1 EPI loci occluded layer o 2 (b) 0 base layer d 1 0 s e d 2 d 3 supporing-surface relief-surface side layer s 2 occluded layer o 1 occluding objec d 4 Fig. 5. A 1D illusraive scene and is LAMP represenaion. (a) 1D scene wih four horizonal dephs (d 1 -d 4 ) and wo sides (s 1,s 2 ), (b) relief-like LAMP and deph layering for he image-based LAMP from an EPI. Noe ha in order o represen all he informaion presened in an image sequence of he scene, differen parallel viewing direcions in (a) are needed, which are refleced in he LAMP represenaion as muliple layers in (b). 1). I is a panoramic image-based represenaion wih 3D informaion. A large-scale scene will firs be represened by a base layer, which is indexed by a seamless 2D panoramic view image (PVI) consising of boh exure and deph maps. This 2D PVI (see Fig. 4a for a real example) is defined as he supporing surface (back surface) of he base layer in he relief-like LAMP, which makes he LAMP represenaion very efficien for archiving (modeling) and rerieval (rendering). 2). I has adapive image resoluion. In he base layer (which includes dephs d 1, d 2, d 3 and d 4 in Fig. 5b), each poin has an aached sreak of muliple pixels in he x direcion, wih boh exure and deph informaion, in order o represen he resoluion loss in he muli-perspecive PVI (he supporing surface). The slope of he locus, i.e. he image velociy v in Eq. (3), gives us he 11

13 number of pixels if i is greaer han 1 (v>1). Oherwise he number of pixels is 1. Real examples can be found in he righ hand side of he panoramic image in Fig. 4a where he doors of he buildings and he bushes in fron of i appear narrower han he real ones. In he LAMP represenaion, he number of pixels ( emporal resoluion ) a each poin in he supporing surface (where x = 0) adapively changes wih he deph of ha poin. The adapive emporal sampling in he horizonal direcion and he inheren adapive spaio-sampling of perspecive projecions in he verical direcion recover and preserve image resoluions of he original frames in a saisfying way. The name relief-like comes from he fac ha he appearance of he fron surface (relief surface) of his represenaion is somewha like a relief sculpure, in which forms and figures (wih image velociies v > 1) are disinguished from a surrounding planar surface (he supporing surface) (Fig. 5b). Fig. 4b shows he corresponding relief surface of he base layer of he relieflike LAMP represenaion for he MB sequence (refer o Fig. 14 for all he inernal daa). Each pixel in he relief-like LAMP is associaed wih a locaion wih (x,y,) coordinaes; for he supporing surface, we have x = 0. The end poin in he relief surface in he locaion (y,) is exacly conneced wih he sar poin in he supporing surface in he locaion (y, +1) (Fig. 5). This feaure allows us o generae seamless mosaics in he image-based LAMP represenaion in he nex secion. 3). I is a layered represenaion. Addiional occluded layers, which are smaller pieces of muliperspecive view images compared wih he complee based layer, represen he occluded and side regions (regions o 1, o 2, s 1 and s 2 in Fig. 5b) ha are no visible in he seleced parallel panoramic view of he base layer (which is orhogonal o he camera pah in Fig. 5). They are aached o he base layer wih he same represenaion (exure, deph and adapive resoluion) as he base layer. Each of he occluded x-segmens is aached o a deph boundary poin (y, 0 ), and has an x coordinae, a sar ime s and an end ime e ha mark is posiion in he xy image (Fig. 5b). We wan o emphasize ha in order o represen all he informaion presened in an image sequence of he scene, differen parallel viewing direcions (shown in Fig. 5a) are needed, which are refleced in he LAMP represenaion (Fig. 5b) as muliple layers wih differen x coordinaes. 12

14 4). I is a muliple perspecive image. The PVI is a muli-viewpoin perspecive image (i.e., he y- image). Each sub-image (a one-column sli-image in concep) in he muli-perspecive panorama is full perspecive, bu successive sli images have differen viewpoins. The muliperspecive represenaion acs as a bridge in he image-based modeling and rendering pipeline beween he modeling end from he original perspecive sequences and he rendering end for new perspecive images, boh wih changing viewpoins over a large disance. In conclusion, a relief-like LAMP is composed of a complee base layer and a se of occluded layer pieces. I can be viewed as an essenial par of he xy image cube in which each pixel has wo aribues - exure (represened by inensiy I ) and deph (represened by he image velociy v ) - conneced o is coordinaes (x,y,), where he x coordinae is implicily represened. The daa srucure of a relief-like LAMP is a spaio-emporal (ST) 2D array ha is defined as follows: sruc x-sreak { in x 0 ; // sar x coordinae, indicaing he viewing direcion of he supporing surface in xlen; // lengh of he x-sreak, indicae hickness of he relief in I[xlen]; // exure: inensiy array of he x-sreak floa v[xlen]; // deph: image velociy array of he x-sreak } sruc Occlusion-Segmen { char ype; // segmen ype: OCCLUDED or SIDE in s, e ; // sar frame and end frame of he segmen x-sreak r[ e - s +1]; // x-sreaks in his segmen } sruc Relief-LAMP-Elemen { x-sreak *r; // a x-sreak Occlusion-Segmen *occ; // a occluding segmen; NULL if no a deph boundary } Relief-LAMP-Elemen rlamp[ydim][tdim]; // rlamp[y][]: y=0~ YDIM, = 0~TDIM I is clear ha a relief-like LAMP is a 2D spaio-emporal array indexed by y and ; a each (y,) locaion here is an x-sreak, which is a 1D segmen along he x direcion inside he xy cube and an opional Occlusion_Segmen ha consiss of a number of x-sreaks. In he curren implemenaion, each x-sreak assigns he same image velociy v of he sar poin (x 0,y,) in he supporing surface. In oher words, for resoluion enhancemen, we only sample v pixels along he x direcion and assign hem he same v value. An Occlusion-Segmen is one of wo ypes - OCCLUDED ha has he same deph, or SIDE ha has linearly inerpolaed dephs. Clearly, he 13

15 represenaion here allows furher refinemen of he deph map in ha each (x,y,) pixel in he relief-like LAMP can has is own deph. The 3D coordinaes of each pixel (x,y,) in he relieflike LAMP can be recovered by he following equaion Z V V V = F, Y = y, X = x V (4) v v v given v(x,y,) in he relief-like LAMP represenaion. 4. IMAGE-BASED 3D LAMP REPRESENTATION A relief-like LAMP can be viewed as having adapive ime sampling for every single pixel in he panoramic supporing surface as well as including all he occluded regions represened in he same way. This represenaion is good for a complex scene ha has many small deph layers. However, in erms of daa srucures, he relief-like LAMP is an inhomogeneous represenaion raher han a se of homogeneous 2D image arrays. In addiion, he base layer usually includes objecs a differen deph levels, and he occluded layers are no merged ino he regions hey belong o. For example, regions o 1 and o 2 in Fig. 5 belong o he same deph surface as region d 3, bu hey are segmened ino wo addiional occluded layers. Hence a naural exension of he basic LAMP is o exrac from i a more concise represenaion - an image-based 3D LAMP. In an image-based 3D LAMP, a scene is represened as muliple layers of 2D mosaiced images in which each layer is ruly represened by wo 2D arrays - a exure map and a deph map. We wish o creae a panoramic mosaic image for each layer ha is boh seamless and preserves adequae image resoluions. There are wo seps necessary o fulfill his goal: deph layering and ime re-sampling Deph Layering The moivaion for layering is o represen occluded regions and differen spaial resoluions of objecs wih differen deph ranges in differen layers. An image based LAMP is layered according o occluding relaions raher han merely dephs, which is also used by oher researchers as a powerful observaion o generae a more compac image-based represenaion [26]. The scene pars wih varying (bu no disconinuous) dephs in a single layer will furher use adapive emporal sampling raes in he second sep o represen differen resoluions for 14

16 differen dephs along he direcion of parallel projecion. Concepually, he implemenaion of deph layering from a relief-like LAMP is sraighforward since we can easily deermine where he occluding regions in he relief-like base layer should be segmened and pu ino separae image-based layers. The same is rue for he occluded or side regions in he occluded layers. Generally speaking, layers eiher in he same deph range or wih coninuous dephs will be merged ino one single layer. Regions wih occluding boundaries will be divided. For example, in Fig. 5b, he wo occluded regions ( o 1 and o 2 ) will be merged ino he base layer wih deph d 3, while he occluding region (d 4 ) originally included in he base layer will be separaed ou as a new layer. Ideally, side regions ( s 1 and s 2 in Fig. 5b) should be insered ino he base layer since hey can connec well (in boh deph and exure) wih he base layer. However, in our curren implemenaion we pu hem ino wo separae layers for simpliciy. Afer deph layering, he original relief-like LAMP is divided ino several single-layer relief-like LAMPs, ready for creaing image-based layers. Each of hem only includes a base layer, wih an x coordinae o indicae where i is aken from in he xy cube. Differen x coordinaes in he relief-like LAMP imply differen viewing direcions of parallel projecions in order o beer capure surfaces wih differen orienaions (see Fig. 5a ). +1 x-segmen +1 Supporing-surface (back) w re-sampling Image-based LAMP Relief surface (fron) Relief-like LAMP y x Fig. 6. Temporal re-sampling and seamless mosaicing 4.2. Temporal Re-Sampling From each single-layer relief-like LAMP ha is obained by deph layering (as shown in he lef par of Fig. 6), we will generae a seamless 2D panoramic image wih boh exure and deph aribues. In each column () of such a relief-like LAMP (Fig. 6), he number of pixels in he x direcion of a poin in he supporing surface is inversely proporional o he deph of ha poin. Since dephs may change along he y direcion, he x-y slice a each column usually has varying widhs in he x direcion. In order o creae a regular 2D image, we wan o warp each irregular x- 15

17 y slice (in he relief-like LAMP) wih varying widhs in he x direcion ino a recangular -y slice of equal widh w (w 1) in he direcion, which will be siched ino a 2D seamless mosaic in he image-based LAMP (see he righ par of Fig. 6). The widh of he -y slice image a ime is calculaed as he dominan image velociy of all v(y, ) in column of he supporing surface, e.g. w = v( ) median{ v( y, )} (5) = y Noe ha each x-segmen of he x-y slice in a relief-like LAMP sars from x 0 and ends a x e (y,)= x 0 -v(y, ); for he base layer x 0 =0. A emporal re-sampling is performed for each x-segmen by urning i ino a -segmen of w -pixels (Fig. 6 and Eq. (5)). In his way, super-ime sampling is virually achieved wihin a frame s ime (usually 1/30 second as he uni) such ha each pixel of a ransformed w -pixel slice represens a poin capured in a ime uni of 1/w. Hence, he exure map and he deph map of a layer saring from ime 0 are represened as I(y,k) and v(y,k) where index k for ime is k [ 1 τ 1 v( τ ), v( τ ) + v( )], v( ) = 0, > (6) = 0 τ = For compuing 3D coordinaes, he super-sampled k corresponding o column k is sored in a 1D array as par of he image-based LAMP model: 1 k v( τ ) τ = 0 k = + [, + 1), v( 0) = 0, > 0 v( ) (7) In his manner, each column k in he image-based LAMP array is virually a 1-column perspecive image a he super viewpoin (ime) k. The densiies of viewpoins adapively change wih dephs of he scene. The parallel-perspecive views beween ime and +1 are approximaed by ransforming he corresponding par of he perspecive image in frame. Noe ha his is a correspondenceless approach o implemening view inerpolaion and is differen from he global parameric mehod in [24] or he local mach mehod in [16, 27] for image mosaicing. Fig. 7 shows a seamless adapive-resoluion panoramic mosaic (corresponding o he PVI shown in Fig. 4a), where he ime scale in each insan is deermined by he dominan 16

18 (median) deph of poins along he corresponding verical line (he y direcion) in frame. Noe ha he aspec raios of he objecs in he righ-hand side of Fig. 4a are resored. rees Fig. 7. Muli-viewpoin mosaic wih adapive ime scales (he righ edge of he upper par connecs wih he lef edge of he boom par). The widh of each verical slice from he corresponding original frame is deermined by he dominan image velociy v of pixels along he y-axis in he corresponding PVI (he supporing surface of he base layer). Circles in his figure indicae he corresponding OCCLUDED and SIDE regions ha are shown on he EPIs in Fig. 12. In conclusion, in an image-based 3D LAMP each layer is basically a 2D parallel-perspecive panorama wih an x coordinae (o indicae viewing direcion). I has hree componens: 1) exure map I(y,k); 2) deph map v(y,k); and 3) a 1D super ime-sampling array k =(k) (o indicae densiies of viewpoins, or adapive emporal resoluions). The daa srucure is sruc Image-LAMP-Layer { in I[YDIM][KDIM]; // exure map I[y][]: y = 0~YDIM; k = 0 ~KDIM floa v[ydim][kdim]; //deph map v[y][] : y = 0~YDIM; k = 0 ~KDIM floa [KDIM]; // super-ime index k ; k = 0~KDIM; } Image-LAMP-Layer ilamp[l]; // layer l = 0~L From his represenaion, we can find he corresponding 3D coordinaes of a poin (y,k) in he image-based LAMP by Z V V V = F, Y = y, X = x V k (8) v v v given v(y,k) and k = (k). 17

19 As a comparison, Table 1 shows he radeoff beween relief-like and image-based LAMP represenaion. While he former is more general and more appropriae for complex scenes, he laer is more compac. Table 1. Comparison beween relief-like LAMP and image-based LAMP represenaions Relief-like LAMP Image-based LAMP Scene general, complex scenes well-layered scenes, 2-3 layers Represenaion inhomogeneous homogeneous, more compac Consrucion inermediae level higher level Rendering fas faser 5. DEPTH, OCCLUSION AND RESOLUTION RECOVERY In his secion, we will discuss one approach our panoramic EPI approach [15, 21] - o obain a dense deph map for he supporing surface of he relief-like LAMP. Furher, we will show how o generae occluded layers and adapive resoluion by selecively using he mos essenial informaion from all he EPIs. The suppor surface is a PVI wih seleced parallel viewing direcion; for example, a viewing direcion orhogonal o he camera moion direcion when x=0. The panoramic EPI approach consiss of four imporan seps: locus orienaion deecion, moion/deph boundary localizaion, deph-exure fusion, and occlusion/resoluion recovery. The resuls of he firs hree seps are a panoramic deph map wih accurae deph boundaries as well as he corresponding exure map (Fig. 4a) Deph Recovery: he Panoramic EPI-Based Approach The panoramic EPI approach we proposed in [15, 21] uses a frequency-spaio-emporal crossanalysis o esimae loci s orienaions in each EPI, wihou explicily racking he loci in EPIs. We will give a brief summary of how his goal is achieved in he firs hree seps of he panoramic EPI approach. The mehod consiss of hree seps: frequency domain loci orienaion esimaion, spaio-emporal domain deph boundary localizaion, and deph-exure fusion. Since i is very imporan o accuraely localize he deph boundary for layered represenaion in imagebased rendering applicaions, we will focus on he mehod for accurae deph boundary localizaion. 18

20 Firs, a Gaussian-Fourier Orienaion Deecor (GFOD) is performed along a scanline (x=x 0 ) in he EPI, which is he inersecion line of his EPI wih he supporing PVI. The GFOD operaor uses Gaussian-windowed Fourier ransforms o deec orienaions of he image under he Gaussian window. A large window (e.g ) is used in order o deec accurae locus orienaions. Muliple orienaions are deeced for a cerain emporal range when he GFOD operaor moves across a deph boundary. Thus, he Gaussian window is applied o reduce his range by assigning higher weighs for pixels closer o he cener of he window. However, he response of muliple (wo in our curren implemenaion) orienaions does no only happen exacly a he poin on he deph boundary (see Fig. 8 for locaions wih wo peaks). x (a) θ (b) θ (c) y (d) Fig. 8. Muliple orienaion deecion by Gaussian-windowed Fourier Orienaion Deecor (GFOD), and deph boundary localizaion. (a). An x- image (i.e., EPI, wih y = -18) wih he processed poins (whie dos) and a Gaussian window (indicaed by a circle). (b) Orienaion energy disribuion map Pd(φ,) and loci orienaion deecion: he long dashed curve (red in color version) indicaes he firs seleced peaks, and several pieces of solid lines (blue in color version) indicae he second peaks. (c) Loci orienaion angles seleced afer deph boundary localizaion and deph inerpolaion. (d) Par of he corresponding PVI (x = 0); he horizonal line in he PVI corresponds o he EPI in (a). Significan deph boundaries are marked by verical black lines across (a) o (d). The Fourier ransform G ( ξ, ω) is obained for a Gaussian-windowed EPI paern cenered a (x 0, ) (shown as a circle in Fig. 8a). The energy specrum P ( ξ, ω) =log(1+g 2 ( ξ, ω) ) is mapped ino he polar coordinae sysem ( r,φ ) by a coordinae ransformaion 19

21 2 2 π ξ r = ξ + ω, φ = + arcan. From he resuling polar represenaion P ( r,φ ), an orienaion 2 ω hisogram is consruced as ( r, φ ) drφ [ 0 π ] P ( φ ) 2 P, (9) d = r r1 where φ corresponds o he orienaion angle of he ST exure cenered a (x 0, ) and [r 1,r 2 ] is a frequency range of he bandpass filer, which is seleced adapively according o he spaialemporal resoluion of he image. Iniially, r 1 and r 2 are se o 8 and 30, respecively, for a window. An orienaion energy disribuion map P d (φ,) can be consruced (Fig. 8b), which visually represens he dephs of he poins along he ime () axis, corresponding o he processed EPI. Fig. 9. Panoramic deph consrucion: raw deph map, refined deph map and he exure maps wih deph boundaries superimposed in black lines (red in color version). In he second sep, a Deph Boundary Localizer (DBL) is used o accuraely localize he deph boundary. I measures inensiy consisencies along he wo deeced orienaions wih he wo highes peaks, aking occlusion/reappearance relaions ino accoun. The bes consisen measuremen should be achieved righ a he deph boundary since oherwise one of he measuremens will cross boh locus paerns (Fig. 8 c; see also Fig. 10a and Fig. 10b). Then in he hird sep, a deph-exure fusion (DTF) algorihm is applied o reduce he errors produced in each EPI analysis (due o aperure problems, noise, ec.) and o refine deph 20

22 boundary localizaions. The refinemen is based on he observaion ha a deph boundary almos always coincides wih an inensiy boundary in a visual scene. This observaion is also used by ohers [22, 25]. Fig. 9 compares he raw deph map (afer he firs wo seps) and he refined deph map, and shows superimposed deph boundaries in he panoramic exure map Deph Boundary Localizaion Since muliple orienaions are deeced no only a bu also near he deph/moion boundaries by using he large GFOD operaor, a Deph Boundary Localizer is designed o deermine wheher or no he deph/moion boundary is righ in he cener of he Gaussian window. For he mehod o be valid for mos of he cases encounered in a naural scene and applicable o he EPIs generaed by a un-sabilized camera, we use an approach ha does no rely on locus racking (which ofen fails due o he non-ideal ST exures generaed from a complex scene wih changing illuminaions and un-smooh camera moion). In our algorihm, muliple scale inensiy similariies are measured along all he deeced orienaions θ ( k = 1, L K ) by he GFOD operaor. k, Among hem he orienaion wih he greaes similariy measuremen is seleced as he correc orienaion. Noe ha only a comparison-and-selecion operaion is used, wihou assuming any deecion of feaure poins or using any roublesome hresholds. Consider he case in which wo orienaions θ 1 and θ 2 ( θ 1 > θ2 ) are deeced wihin a Gaussian window. Dissimilariy (i.e., variance) measuremens along θ 1 and θ 2 for a given circular window of radius R cenered a he poin (x 0, 0 ) are defined as he variance of inensiy values (Fig. 10 (a) and (b); refer o Fig. 8) C 1 θ (k=1,2) (10) R ± k k R r= (, R) = I ( ± r, θ ) I ( θ, R) 2 2 where r = ( x x ) + ( ), I (, R) = I( r θ ) ± k 1 θ. R 0 0 ± k ±, k R r= 1 In he above equaions, subscrips + and - denoe he dissimilariy measuremens along he deeced orienaions in posiive (+) and negaive (-) x direcions respecively. This is designed for dealing wih he occlusion of a farher objec ( θ 2 ) by a closer one ( θ 1 ): he occluding (i.e., closer) objec can be seen in boh he posiive and he negaive x direcions, bu he occluded (i.e., farher) objec can only be seen in one of he wo direcions (Fig. 10a). The dissimilariy measuremens for closer and farher objecs are defined as 21

23 E E ( θ, R) = ( C ( θ, R) + C ( θ, R) )/ P ( θ ) 1 ( θ, R) = min( C ( θ, R), C ( θ, R) )/ P ( θ ) d d 1 2 (11) respecively. Noice he difference in he wo measuremens - he occluded objec only akes he smaller measuremen among he posiive and negaive direcions. In addiion, we give more weighs o sronger oriened exure paerns: P d ( θ k ) is he value of he orienaion hisogram (Eq. (9)) a θ k (k=1,2). The higher he value is, he lower he dissimilariy measuremen should be. x C + (θ 1,R) C + (θ 2,R) x C + (θ 1,R) C + (θ 2,R) x C + (θ 1,R) C + (θ 2,R) C - (θ 2,R) (x 0, 0 ) C - (θ 2,R) (x 0, 0 ) C - (θ 2,R) (x 0, 0 ) C - (θ 1,R) C - (θ 1,R) C - (θ 1,R) (a) occlusion: muli-peaks in hree cases o he lef, jus a and o he righ of a deph boundary x C + (θ 1,R) C + (θ 2,R) C - (θ 2,R) (x 0, 0 ) C - (θ 1,R) (x 0, 0 ) x θ 1 θ 2 θ3 R 3 θ 3 R R 2 1 Linear(θ 1, θ 2 ) (b) reappearance (c) muli-scale window (d) deph inerpolaion Fig. 10. Principle of he deph boundary localizaion and deph inerpolaion Fig. 10a shows how o use hese measuremens o localize a deph boundary when he farher objec will be occluded by he closer objec (which is he occlusion case). Muliple peaks are deeced by he GFOD operaor when he Gaussian window (indicaed by circles) is near he deph boundary. When he Gaussian window is o he lef of he deph boundary, he dissimilariy measuremen (i.e. variance) along he locus direcion of he occluding objec E, 1 ( θ R) will be larger, since he measuremen is performed across he loci paern of he o-beoccluded objec (lef of Fig. 10a). On he oher hand, he dissimilariy measuremen along he locus direcion of he o-be-occluded objec E ( θ, R) will be much smaller, since he measuremen is righ along he locus of he o-be-occluded objec. Whenever he cener of he Gaussian window is precisely a he deph boundary, boh measuremens will be small since boh 2 22

24 measuremens are along heir own loci s direcions. However, since he occluding boundary of he closer objec usually will be visually sronger han he ST paern of he occluded objec, he measuremen will be in favor of he closer objec a his locaion (middle of Fig. 10a). As he cener of he window moves ino he occluding (closer) objec, he dissimilariy measuremen of he occluded objec will be significanly increased, since he measuremen will cross he loci of he occluding objec, bu he dissimilariy measuremen for he occluding objec will remain small (righ of Fig. 10a). Similar argumens hold for he reappearance case, when he occluded objec reappears behind he occluding (closer) objec (Fig. 10b). Therefore, a simple verificaion crierion can be expressed as θ1, if E( θ1, R) E( θ2, R) θ = θ 2, Oherwise (12) In fac, he condiion of occlusion and reappearance can be judged eiher by comparing + ( R) and C (, R) C θ 2, θ 2 (see Fig. 10) or by analyzing he conex of he processing (i.e., he change of dephs). In he case of occlusion of a far objec by a near objec (far o near, Fig. 10(a)), we have C ( R) C (, R) ( θ R) C ( R) C + 2, < θ2,. θ 2, + θ2 <, and in reappearance (near o far, Fig. 10(b)) we have In order o handle cases of various objec sizes, differen moion velociies, and muliple objec occlusions, muliple scale dissimilariy measuremens E (, ) θ (e.g., i=1,2,3) are calculaed wihin muliple scale windows of radii R i (i=1,2,3), R 1 <R 2 <R 3. In our experimens, we have seleced R 1 =m/8, R 2 =m/4, R 3 =m/2 (m = 64 is he window size; see Fig. 10(c)). By defining he following raio D max ( E( θ1, Ri ), E( θ2, Ri )) ( E( θ, R ), E( θ, R ) k R i i = (13) min 1 i 2 i a scale p (p=1,2 or 3) wih maximum Dp is seleced for comparing he inensiy similariies. For example, in Fig. 10(c), R 2 will be seleced. The seleced orienaion angle θ can be refined by searching for a minimum dissimilariy measuremen for a small-angle range around θ. The accuracy of he orienaion angle, especially ha of a far objec, can be improved by using more frames. 23

25 In order o obain a dense deph map, inerpolaions are applied o exureless or weak-exured regions/poins where no orienaion can be deeced. The proposed inerpolaion mehod (Fig. 10(d)) is based on he fac ha a deph disconinuiy almos always implies an occluding boundary or shading boundary. The value θ () beween wo insans of ime 1 and 2 wih esimaed orienaion angles θ 1 and θ 2 is linearly inerpolaed for smooh deph change (i.e., θ, T dis is a hreshold), and is seleced as min( θ 1, θ2 ), i.e., he angle of he farher 1 θ2 < Tdis objec, for deph disconinuiy (i.e., θ 1 θ2 Tdis measuremens from a real x- image(epi) are shown in Fig. 8c. ). The processing resuls of dense deph 5.3. Occlusion Recovery Because he supporing PVI only conains informaion from a single viewing direcion (for example, he direcion perpendicular o he moion direcion when we selec a PVI wih x 0 = 0), some pars of he scene ha are visible in oher pars of images from an original (or a sabilized) video sequence are missing due o occlusion. They will be recovered by analyzing deph occlusion relaions in he EPIs. The basic algorihm is performed in each EPI afer he panoramic deph map and is deph boundaries have been obained. The algorihm consiss of he following seps (Fig. 11, Fig. 12): Sep 1. Find he locaion of a deph boundary - A poin on a deph boundary, p0(x 0, 0 ), and orienaion angles (θ 2 and θ 1 ) of he occluded (far) and occluding (near) objecs are encoded in he deph map. A poin is considered as a poin on he deph boundary wih a deph disconinuiy when, for example, o θ 1 θ 2 > 2. Sep 2. Localize he missing par - The missing (occluded) par is represened by a 1D (horizonal) spaio-emporal segmen p s p e in he EPI, on which poins have he same parallel viewing direcions bu from moving viewpoins. I is calculaed from he slopes of he wo orienaion paerns ha have generaed he deph boundary, and i is denoed by an x coordinae and sar/end frames (s/e). Basically, he larges possible angle of viewing direcion (indicaed by he x posiion in he EPI) from he viewing direcion of he PVI (indicaed by he x 0 coordinae) possesses he mos missing informaion, bu possible occlusions by oher nearby objecs should be considered. For example, he second missing region from he righ in Fig. 12b was deermined by checking he occlusion of he locus paerns of he missing par agains hose 24

26 of oher nearby foreground objecs (rees), resuling in an ST segmen wih smaller x coordinae, i.e. smaller viewing angle from he viewing direcion of he PVI. In his way, a riangular region p 0 p s p e can be deermined, and he 1D segmen p s p e will be used as he exure of he missing par ha is occluded by he foreground objecs. x p s p e θ 1 θ 2 p 0 x p 0 p s θ 1 θ 2 p e (a) (b) (c) x p 0 p x p 1 (0,+1) (x,) Fig. 11. Region classificaion and adapive resoluion. (a) Locus paerns near an occlusion boundary; (b) locus paerns of fron- side surfaces and (c) emporal resoluion enhancemen by using spaial resoluion. Sep 3. Verify he ype of he missing par. The riangular region also conains deph informaion of he missing par - he 1D segmen p s p e. In principle, similar reamens can be made here as for he basic deph map as in Eq. (10). For simpliciy, he missing pars are classified ino wo ypes in our curren implemenaion: OCCLUDED and SIDE. If he loci wihin he riangular region form a parallel paern of angle θ 2 (Fig. 11a), hen he missing par is classified as OCCLUDED; oherwise, as SIDE if he angle of he oriened paern θ2 θ1 θ = θ1 + ( s) (14) e s changes linearly inside he riangle (Fig. 11b). By assuming he missing par as each of he wo ypes, an overall consisen measuremen along he hypohesized locus orienaions (indicaed by arrows in Fig. 11a and b) can be calculaed wihin he riangle region (similar o measuremens in moion boundary localizaion [15]). The ype of missing par is seleced as he one wih he beer consisen measuremen of he wo hypoheses. The loci s angle θ of he OCCLUDED region will be he same angle θ 2 as he occluded objec, whereas loci s angle θ of he SIDE region gradually changes from θ 1 o θ 2 (or from θ 2 o θ 1 ), as expressed in Eq. (14). 25

27 In his way, he dephs of he occluded or side region can be assigned. Fig. 11 illusraes he siuaion of reappearance where he farher objec re-appears behind he closer objec. Fig. 12 shows oher siuaions (occlusion and side) wih real image.: Fig 12(a) shows an example of recovering an occluded region (building façade) behind a ree, as indicaed by he firs circle in Fig. 7. Fig 12(b) shows hree recovered side regions. The firs wo correspond o he firs side of he building indicaed by he second circle in Fig. 7, which is separaed ino wo regions by a ree in fron of i. The hird side region corresponds o he second side façade indicaed by he hird circle in Fig. 7. The x locaion of he second side region is much closer o he supporing PVI (wih x = 0), since i is occluded by he ree. (a) p 0 (b) rees p 0 p s p e p s p e OCCLUDED SIDE Fig. 12. Occlusion and resoluion recovery resuls in real EPIs. (a) an OCCLUDED region; (b) hree SIDE regions (wo of hem belong o a side facade separaed by rees). Circles in Fig. 7 show he corresponding OCCLUDED and SIDE regions in his figure Resoluion Recovery As we have shown, he panoramic view image (PVI) in he direcion is under-sampled if he image velociy v of a poin in he PVI is greaer han one pixel per frame (as in he righ hand par of Fig. 4a, where he images of doors of he building look hinner han he real ones). Το enhance emporal resoluion, a v-pixel segmen in he x direcion is exraced from he EPI (as opposed o a single pixel in a PVI, as in Fig. 2 and Fig. 4a). Fig. 11c shows he principle: a nice feaure is ha he fron (relief) poin p x (x,) of an x-segmen p 0 p x a ime will exacly connec wih he back (supporing) poin p 1 (0,+1) of he x-segmen a ime +1 in he relief-like LAMP represenaion, since boh image poins are on he same locus of a 3D poin. This propery has been used o generae seamless, adapive-ime panoramas in Fig. 7. The hickness of he red 26

28 (dark in B_W version) horizonal lines (including hose of SIDE and OCCLUDED regions) in he wo EPI iles in Fig. 12 indicaes he number of poins (pixels) o be exraced in he x direcion of his epipolar plane image. As shown in Fig 7, a seamless panoramic mosaic is generaed afer resoluion enhancemen, where much higher emporal resoluions are achieved in he second par of he mosaic. video (1a) video sabilizaion (1b)PVI & EPI generaion Panoramic exure map (2a) Deph Map Acquisiion EPI 1 Loci orienaion & moion boundary EPI 2 esimaion EPI H H: heigh of a frame Panoramic deph map (2b)Texure-deph fusion (3). Occlusion & resoluion recovery Relief-like LAMP Image-based LAMP Image-based rendering Fig. 13. Sysem diagram of 3D panoramic scene modeling (PVI: Panoramic View image; EPI: Epipolar Plane Image; LAMP: Layered, Adapive-resoluion and Muliperspecive Panorama) 6. LAMP CONSTRUCTION AND RENDERING: EXPERIMENTAL RESULTS D LAMP Consrucion Resuls Our 3D panoramic scene modeling approach consiss of hree modules: (1) video sabilizaion (and PVI/EPI generaion), (2) dense deph map acquisiion (wih exure-deph fusion), and (3) deph layer consrucion (wih occlusion/resoluion recovery). The sysem diagram is shown in Fig. 13. For using he EPI-based approach o recover dense deph maps, our mehod requires a dense, long, 1D ranslaional video sequence as he inpu in order o generae parallel-perspecive panoramic mosaics. I is paricularly effecive when he number of frames is larger han he widh of an original video frame in order o generae mosaics wih panoramic view. In he 27

29 ideal case, he represenaion of parallel projecion requires a pure 1D ranslaion along one image axis. However our algorihms sill work under more pracical moion given ha a video sabilizaion module is used o preprocess he sequence. Our 3D video sabilizaion module esimaes inerframe camera moion, models he camera moion of he image sequence as a 1D dominan ranslaion plus a small vibraion (he laer is modeled by a homography beween successive frames), and removes he vibraion by a moion filering and image warping sep so ha a recified video sequence wih virually 1D ranslaional moion can be creaed. Deailed algorihms and resuls can be found in [15, 21]. The deph map acquisiion module was summarized in Secion 5.1. LAMP represenaions and consrucion are presened in Secions 3 and 4, while he algorihms for occlusion/resoluion recovery has been discussed in Secions 5.2 and 5.3. In his secion, we will presen some consrucion resuls from real image sequences, look a some pracical consideraions in LAMP represenaions, and discuss he LAMP-based 3D rending issues. (a) occlusion (b) afer occlusion recovery Fig. 14. Inernal daa of (a) he base layer, and (b) all he layers of (par of) he LAMP of he MB sequence. The images are runcaed for fiing ino he page. The MB sequence was capured by a camcorder mouned on a ripod carried by a moving vehicle. Aside from he small vibraion of he vehicle on a normal oudoor road, he velociy of ranslaion was mosly consan. The small vibraion in he video sequence (which is demonsraed by he irregular boundaries of he mosaic in Fig. 7) was removed by a video sabilizaion process discussed in our previous work[15, 21]. The frequency-domain deph esimaion mehod is robus for dealing wih he non-ideal loci creaed from his real-world video 28

30 sequence wih vibraion. We have shown he supporing surface and he relief-surface of he relief-like LAMP represenaion consruced from he MB sequence; Fig. 14 shows he inernal daa for he base layer and all he layers. Recall ha a relief-like LAMP represenaion is par of a 3D xy cube. Tha is o say, for each pixel (y,) in he supporing surface, here is an x-segmen of v pixels in he x direcion. The inernal daa of he base layer is displayed in Fig. 14a as he sequenial head-ail arrangemen of all he segmens of he base layer in each row. We have found ha boh he exure and he deph map show almos seamless connecions beween successive segmens, and i is obvious ha higher resoluions han in he corresponding PVI are recovered for closer objecs. The inernal daa of he relief-like LAMP, including all he occluded layers, is displayed in Fig. 14b as similar head-ail arrangemens of all he x-segmens in all he layers, including all he occluded layers as well as he base layer. The x-segmens in he occluded layers are insered ino he locaions of heir corresponding deph boundaries. Comparing Fig. 14b o Fig. 14a, addiional daa for he occluded regions are shown. For example, he porion of he building s facade occluded by he rees and he side facades of he building are parially recovered. Noe ha since boh images are runcaed o he same lengh, leaving ou differen amouns of scene poins on he righ side. The compacness of he LAMP represenaion can be seen from he following key numbers for he MB sequence (W*H*F= widh*heigh*frames, P: number of pixel per locaion in he PVI): - Original video sequence: S 0 = W*H*F=128*128*1204 byes (gray level images) = 16MB - PVI supporing surface (boh exure and deph in 1 bye/pixel): S p =H*F*2 = 128*1024*2 byes = 256KB - LAMP represenaion (exure 1 bye/pixel, deph 4 bye/pixel, as in he x_sreak srucure in Secion 3): MB. The las number approximaely corresponds o P = 3 pixels per locaion and 5 byes per pixel on average in he PVI supporing surface, which coun for boh adapive resoluion and occlusion represenaion. This ends up wih an esimaed LAMP size of 128*1024*5*3 = 1.92MB, which is close o he real size. If we use 1 bye/pixel o represen deph, hen he LAMP size will be S l = H*F*2*P = 128*1024*2*3 = 768 KB, which is simply hree imes ha of he PVI supporing surface. 29

31 In conclusion, he size of he LAMP represenaion in his example is abou 1/10 (if floaing poin deph is saved) or 1/20 (if bye/pixel deph is saved) of he daa size of he original video sequence. In he general case, he LAMP-o-sequence size raio is S l /S o = 2P/W, where W is he widh of each frame in he original video sequence and P is he number of pixels per locaion of he PVI supporing surface Exended Panoramic Image (XPI) Represenaion We have shown in [15] how o make full use of he original image sequence by generaing an exended panoramic image (XPI). Suppose ha an image sequence has F frames of images of size W H. An example is he frequenly used flower garden (FG) sequence(w H F = ). A PVI and an EPI is shown in Fig. 15a and Fig. 15b. I is unforunae in his case ha he field of view of he panoramic view image urns ou o be narrow due o he small number of frames and large inerframe displacemens. Therefore, an exended panoramic image (XPI) is consruced. The XPI is composed of he lef half of frame m/2 (m is he GFOD window size), he PVI par formed by exracing cener verical lines from frame m/2 o frame F-m/2, and he righ half of frame F-m/2 (Fig. 15c). The oal widh of he XPI is W x =W/2 +(F m)+ W/2 = W+F-m. PVI XPI y EPI analysis along his zigzag line EPI occlusion x mxm a b c Fig. 15. PVI, EPI and XPI of he flower garden (FG) sequence Fig. 16 shows he resuls of 3D recovery of he XPI of he FG sequence. In he deph map, he ree runk sands ou disincly from he background, and he gradual deph changes of he 30

32 flowerbed are deeced. The occluded regions and resoluions are recovered in a way similar o ha for a PVI image, excep ha he EPI analysis is performed along a zigzag line (Fig. 15b). a b Fig. 16. Panoramic deph map for he FG sequence. (a) Isomeric deph lines overlaid in he inensiy map (b)panoramic deph map. x-y par -y par (a) background layer d 0.5 k x-y par (b) objec layer Fig. 17. Image-based LAMP model of he FG sequence Fig. 17 shows he wo exraced layers of he image-based LAMP represenaion for he FG sequence, each of which has boh exure and deph maps. These wo pairs of images, along wih a 1D emporal sampling rae array (Eq. (7)) for each layer, are consruced from he 115-frame FG image sequence. In he y par of he background layer of his exended image-based LAMP, he ime-sampling rae is adapively changed according o he dominan deph in each ime insan. The ime scales are less han 0.5 (frames); his means ha higher resoluion is achieved han ha of he original PVI (shown in Fig. 16) in boh exure and deph maps (he uneven 31

33 super-iming in Fig. 17a is due o quanizaion). Furhermore, he background regions occluded by he ree runk have been compleely recovered, and have been merged ino he background layer wih boh exure and deph values. The size of he XPI-based LAMP represenaion in he general case is S xl = (W+(F-m)*P)*H*2*L (byes). where W*H is he size of an original frame, F is he number of frames in he sequence, m is he size of he Gaussian window (m = 64 in our experimens), P is he number of pixels per locaion in he PVI par of he XPI image, and L is he oal number of layers. The real sizes of he imagebased represenaion of he FG sequence is S xl = (114*2 + 95*2) KB = 418 KB. Compared o he size of he original video W H F = = 9.265MB, he real compression raio is The esimaed size using he above equaion is S xl = (352+(115-64)*2)*240*2*2 = 423 KB assuming P = 2 and L = 2. The heoreical LAMP-o-sequence raio is S xl /S o = 2L/F + 2PL/W, assuming ha F is much larger han m LAMP-Based Rendering The 3D LAMP model is capable of synhesizing images from new viewpoins differing from he viewpoins of he original image sequence hanks o is muli-perspecive represenaion, adapive resoluion, occlusion represenaion, and deph informaion. The mapping from a pixel in a relief-like LAMP model o is 3D coordinae can be compued by using Eq. (4), and hen i is sraighforward o re-projec i o he desired view (in pracice an inverse mapping should be applied). Because of he neighborhood relaions of pixels in he LAMP represenaion (refer o Fig. 14), a rendering algorihm can easily perform inerpolaion beween neighborhood pixels. In a relief-like LAMP, an aached layer is always occluded by he layer o which i is aached. So, when rendering he aached layers should be drawn firs so ha an occluded region could be seen correcly in a new view. Thanks o he adapive resoluion represenaion, he higher resoluions in he original images can be applied o a new view whose viewpoin is close o hose of he original image sequence. Fig. 18 shows he preliminary rendering resuls from he relief-like LAMP of he MB sequence. The developmen of a rendering sysem based on relieflike represenaion is sill underway. 32

34 (a) (b) Fig. 18. Rendering resuls from relief-based LAMP represenaion of he MB sequence. (a) Normal rendering (wih correc occlusion checking of differen layers), where he building façade is occluded by he ree. (b) The par of he façade is seen hrough he ree. Noe ha ha par is in shadow in he original video sequence, and herefore i is darker. The rendering process is easier using an image-based LAMP model. For a LAMP model based on an XPI (i.e., a combinaion of x-y and y- images) in general, Eq. (4) and Eq. (8) should be used in he wo x-y pars and he one y- par, respecively. Forunaely, boh equaions are very simple and virually he same (bu reflec he differen ways o obain x and indices) so ha a fas implemenaion is feasible. In order o achieve a correc occlusion relaion, a rendering from a new view should begin wih he farhes layer and end wih he neares layer, from he viewpoin of a synheic image. A simple image based 3D rendering program has been developed o demonsrae he capabiliy of image synhesis of arbirary views using he image-based LAMP represenaion. Synheic sequences wih a virual camera of 6 DOF of moion generaed from he LAMP model (wih and wihou he objec layer) of he flower garden sequence can be found on our web page [20]. The compacness of he image-based LAMP represenaion enables he rendering algorihm o achieve real-ime performance. Fig. 19 shows some snapshos of he synhesis wih boh he background and foreground layers, wihou he foreground layers. The perspecive disorion shown a he borders of he synheic images in he second and he hird rows in Fig. 19 clearly reveals he accuracy in he recovered deph changes for he ree, house, and garden. The firs synheic images in boh he second and he hird rows were generaed from a viewpoin farher from he original camera pah so ha he enire scene can be seen. 33

35 Fig. 19. Synheic images from he image-based LAMP model. Firs row: Snapshos of original video sequence; second row: snapshos of he virual walk-hrough wih boh he background and he rees; hird row: snapshos of synheic resuls wihou he ree. The firs synheic images in he 2nd and 3rd rows are generaed from viewpoin ha can see all he scene poins in he LAMP model. 7. COMPARISONS AND DISCUSSIONS 7.1. Layered Represenaion The LAMP represenaion is relaed o such represenaions as muli-perspecive panoramic view images (PVIs), spries, and layered deph images (LDIs). However, i is more han a muliperspecive PVI in ha deph, adapive-resoluion and occlusion are added. I is differen from a "sprie" (or LDI) since he sprie or LDI is a view of a scene from a single inpu camera view and is wihou adapive image resoluion. We compare our resuls [17] wih oher layered represenaions [10, 11, 13, 22, 23]. A comparison in number of inpu images, resuls of layered represenaions, and algorihm performance for he flower garden sequence is summarized in Table 2. Usually, in a layered represenaion a se of deph surfaces is firs esimaed from an image sequence of a moving camera and hen combined o generae a new view. The layered represenaion proposed in [10] consiss of hree maps in each layer: a mosaicing inensiy map, 34

36 an alpha map, and a velociy map. The velociy map is acually a se of parameers of he affine ransformaion beween layers. Occlusion boundaries are represened as disconinuiies in a layer's alpha map (opaciy). In [10], he firs 30 frames of he flower garden sequence were used as inpu. The oupu were hree affine plane layers - he ree, he house, and he flowerbed. Occlusion (by he ree) was recovered, bu no deph informaion was provided (i.e., each layer is only modeled as an affine plane). Processing of he 30 frames of 720*480 images ook 40 minues in a HP 9000 series 700 worksaion. Table 2. Comparison of 4 layered represenaion resuls for he flower garden sequence Wang-Adelson [10] Ke-Kanade [22] Sawhney-Ayer Baker-Szeliski- Image-based LAMP [11] Anandan[12] Inpu firs 30 frames (720x480) wo frames a few frames firs 9 even frames all 115 frames (350x240) Rep. 3 planar layers (ree, flower bed, house) 4 planar layers (ree, branch, 4 layers (ree, flower bed, 6 layers of Spries (3 ree, 2 flower 2 layers of LAMP (ree, background ) house, flowerbed) house, sky ) bed, 1 house ) Deph no no no yes yes, each obained from 64 frames Occlusion recovered no recovered no recovered recovered recovered Muli-view affine mosaic rep. 2D layer rep. single view single view mosaic muliple-view mosaic mosaic Adapive res. no no no a perspecive image adapive resoluion Performance 40 mins / 30 frames in a HP 9000 series 700 worksaion no available no available no available 14 mins / 115 frames in a 400 MHz PC (general C code) The subspace approach [22] is an effecive mehod for 2D layer exracion from an uncalibraed image sequence. For he flower garden sequence, four planar layers, which roughly correspond o ree, branch, house, and flowerbed, are exraced from wo frames in he sequence. Deph boundaries are accurae afer he layer refinemen sep using color segmenaion resuls (similar o our deph-exure fusion sep). However, deph informaion of each layer is no obained since hey are 2D layers, and he occluded regions are no recovered since only wo frames are used. The muliple moion esimaion mehod based on MDL and EM algorihms in [11] is compuaionally expensive. I requires a combinaorial search o deermine he correc number of layers and he "projecive deph" of each poin in a layer. For he flower garden sequence, only a few frames were processed, and a six parameric moion model was used o segmen he scene 35

37 ino four layers - ree, house, flowerbed and sky. Occlusion regions are no recovered in he layered model, and deph was no esimaed in he flower garden example. In he sprie represenaion [12], each layer consiss of an explici 3D plane equaion, a exure map (a sprie), and a map wih deph offse relaive o he plane. For he flower garden sequence, eigh-parameer homography was used o fi he planar layers. The firs nine even frames were used in he experimen, and he resul was six layers of spries - hree for he ree, wo for he flowerbed, and one for he house. Deph informaion was aached o each pixel, and he occlusion (by he ree) was recovered. Each layer was a single perspecive view mosaic. More recen work on 3D layer exracion by he same research group [23] uses an inegraed Bayesian approach o auomaically deermine he number of layers and he assignmen of individual pixels o layers. The approach is a very general one. However he segmenaion resuls for he flower garden sequence are no as saisfacory as he resuls in [12]. As he auhors poined ou in heir paper, here was always a danger in choosing he prior in order o obain a desired resul of layering, and his remains a challenging research issue. For he flower garden sequence, our mehod segmens he scene ino wo complee layers of he image-based LAMP represenaion: he ree and he background. Each layer has boh a exure map and a dense deph map. The segmenaion is very accurae along he deph boundaries, he occluded regions in he seleced panoramic view are recovered from oher views, and he images have adapive resoluions. Our algorihms can process 115 frames of 350*240 images in 14 minues on a 400 MHz PC o obain he final LAMP model. The processing ime reduces o 2.6 minues on a Xeon 2.4 GHz dual-cpu Dell Linux worksaion. The ime includes reading all of he 240 EPIs from a SCSI hard disk and displaying he processed EPIs and resuling images on he screen Full Perspecive, Full Orhogonal, Muli-Perspecive and Mulivalued Represenaions There are a leas hree exising geomeric represenaions for large image sequences, based on parallel projecion (i.e., a exured digial elevaion map), single-view perspecive mosaicing (e.g. sprie), and parallel-perspecive panorama (i.e., PVI). A parallel projecion of he panoramic deph in Fig. 4a of he MB sequence is shown in Fig. 20a ha correcly shows he aspec raio of deph and size (heigh and widh) of he scene. However, he resoluions of he nearer objecs are significanly decreased (or los) because of his evenly sampled represenaion. 36

38 A single view perspecive represenaion, on he oher hand, is no a good represenaion mehod for surfaces of varying orienaions and over a large disance (Fig. 20b). A very small deph esimaion error in he single-view-based represenaion could be grealy enlarged when reprojecing o a mosaiced image of a single referenced viewpoin. (a) (b) Fig. 20. Two geomeric represenaions of he 3D panorama of he MB sequence. (a) Parallel projecion (b) Perspecive projecion from a viewpoin a he cener of he pah where he video was capured A parallel-perspecive panoramic view image (PVI) is a compac image-based represenaion ha beer represens he image daa capured by a camera ranslaing over a long disance. Based on he muli-perspecive panorama, we have proposed a sill more compac and more powerful represenaion - 3D layered, adapive-resoluion and muli-perspecive panorama (LAMP). To our knowledge, his seems o be he firs piece of work ha inegraes muli-perspecive panoramas and layered represenaions wih adapive image resoluions in a unified model. The relief-like LAMP is basically a single exended muli-perspecive panoramic view image (PVI) wih boh exure and deph values, bu each pixel has muliple values o represen resuls of occlusion recovery and adapive resoluion enhancemen. The image-based LAMP, on he oher hand, consiss of a se of muli-perspecive layers, each of which has boh exure and deph maps wih adapive densiies of viewpoins depending on dephs of scene poins. No assumpion is made on he srucures of a scene in consrucing is 3D LAMP represenaions. The LAMP represenaions are effecive for image-based rendering. We have noiced ha our LAMP represenaions are very similar o he muli-valued represenaion (MVR) proposed by Chang and Zakhor [26] in ha (1) dense dephs are 37

CAMERA CALIBRATION BY REGISTRATION STEREO RECONSTRUCTION TO 3D MODEL

CAMERA CALIBRATION BY REGISTRATION STEREO RECONSTRUCTION TO 3D MODEL CAMERA CALIBRATION BY REGISTRATION STEREO RECONSTRUCTION TO 3D MODEL Klečka Jan Docoral Degree Programme (1), FEEC BUT E-mail: xkleck01@sud.feec.vubr.cz Supervised by: Horák Karel E-mail: horak@feec.vubr.cz

More information

Implementing Ray Casting in Tetrahedral Meshes with Programmable Graphics Hardware (Technical Report)

Implementing Ray Casting in Tetrahedral Meshes with Programmable Graphics Hardware (Technical Report) Implemening Ray Casing in Terahedral Meshes wih Programmable Graphics Hardware (Technical Repor) Marin Kraus, Thomas Erl March 28, 2002 1 Inroducion Alhough cell-projecion, e.g., [3, 2], and resampling,

More information

A Matching Algorithm for Content-Based Image Retrieval

A Matching Algorithm for Content-Based Image Retrieval A Maching Algorihm for Conen-Based Image Rerieval Sue J. Cho Deparmen of Compuer Science Seoul Naional Universiy Seoul, Korea Absrac Conen-based image rerieval sysem rerieves an image from a daabase using

More information

STEREO PLANE MATCHING TECHNIQUE

STEREO PLANE MATCHING TECHNIQUE STEREO PLANE MATCHING TECHNIQUE Commission III KEY WORDS: Sereo Maching, Surface Modeling, Projecive Transformaion, Homography ABSTRACT: This paper presens a new ype of sereo maching algorihm called Sereo

More information

4.1 3D GEOMETRIC TRANSFORMATIONS

4.1 3D GEOMETRIC TRANSFORMATIONS MODULE IV MCA - 3 COMPUTER GRAPHICS ADMN 29- Dep. of Compuer Science And Applicaions, SJCET, Palai 94 4. 3D GEOMETRIC TRANSFORMATIONS Mehods for geomeric ransformaions and objec modeling in hree dimensions

More information

Image segmentation. Motivation. Objective. Definitions. A classification of segmentation techniques. Assumptions for thresholding

Image segmentation. Motivation. Objective. Definitions. A classification of segmentation techniques. Assumptions for thresholding Moivaion Image segmenaion Which pixels belong o he same objec in an image/video sequence? (spaial segmenaion) Which frames belong o he same video sho? (emporal segmenaion) Which frames belong o he same

More information

CENG 477 Introduction to Computer Graphics. Modeling Transformations

CENG 477 Introduction to Computer Graphics. Modeling Transformations CENG 477 Inroducion o Compuer Graphics Modeling Transformaions Modeling Transformaions Model coordinaes o World coordinaes: Model coordinaes: All shapes wih heir local coordinaes and sies. world World

More information

EECS 487: Interactive Computer Graphics

EECS 487: Interactive Computer Graphics EECS 487: Ineracive Compuer Graphics Lecure 7: B-splines curves Raional Bézier and NURBS Cubic Splines A represenaion of cubic spline consiss of: four conrol poins (why four?) hese are compleely user specified

More information

Probabilistic Detection and Tracking of Motion Discontinuities

Probabilistic Detection and Tracking of Motion Discontinuities Probabilisic Deecion and Tracking of Moion Disconinuiies Michael J. Black David J. Flee Xerox Palo Alo Research Cener 3333 Coyoe Hill Road Palo Alo, CA 94304 fblack,fleeg@parc.xerox.com hp://www.parc.xerox.com/fblack,fleeg/

More information

Gauss-Jordan Algorithm

Gauss-Jordan Algorithm Gauss-Jordan Algorihm The Gauss-Jordan algorihm is a sep by sep procedure for solving a sysem of linear equaions which may conain any number of variables and any number of equaions. The algorihm is carried

More information

Video Content Description Using Fuzzy Spatio-Temporal Relations

Video Content Description Using Fuzzy Spatio-Temporal Relations Proceedings of he 4s Hawaii Inernaional Conference on Sysem Sciences - 008 Video Conen Descripion Using Fuzzy Spaio-Temporal Relaions rchana M. Rajurkar *, R.C. Joshi and Sananu Chaudhary 3 Dep of Compuer

More information

Improved TLD Algorithm for Face Tracking

Improved TLD Algorithm for Face Tracking Absrac Improved TLD Algorihm for Face Tracking Huimin Li a, Chaojing Yu b and Jing Chen c Chongqing Universiy of Poss and Telecommunicaions, Chongqing 400065, China a li.huimin666@163.com, b 15023299065@163.com,

More information

Visual Indoor Localization with a Floor-Plan Map

Visual Indoor Localization with a Floor-Plan Map Visual Indoor Localizaion wih a Floor-Plan Map Hang Chu Dep. of ECE Cornell Universiy Ihaca, NY 14850 hc772@cornell.edu Absrac In his repor, a indoor localizaion mehod is presened. The mehod akes firsperson

More information

A Fast Stereo-Based Multi-Person Tracking using an Approximated Likelihood Map for Overlapping Silhouette Templates

A Fast Stereo-Based Multi-Person Tracking using an Approximated Likelihood Map for Overlapping Silhouette Templates A Fas Sereo-Based Muli-Person Tracking using an Approximaed Likelihood Map for Overlapping Silhouee Templaes Junji Saake Jun Miura Deparmen of Compuer Science and Engineering Toyohashi Universiy of Technology

More information

Virtual Recovery of Excavated Archaeological Finds

Virtual Recovery of Excavated Archaeological Finds Virual Recovery of Excavaed Archaeological Finds Jiang Yu ZHENG, Zhong Li ZHANG*, Norihiro ABE Kyushu Insiue of Technology, Iizuka, Fukuoka 820, Japan *Museum of he Terra-Coa Warrlors and Horses, Lin Tong,

More information

In fmri a Dual Echo Time EPI Pulse Sequence Can Induce Sources of Error in Dynamic Magnetic Field Maps

In fmri a Dual Echo Time EPI Pulse Sequence Can Induce Sources of Error in Dynamic Magnetic Field Maps In fmri a Dual Echo Time EPI Pulse Sequence Can Induce Sources of Error in Dynamic Magneic Field Maps A. D. Hahn 1, A. S. Nencka 1 and D. B. Rowe 2,1 1 Medical College of Wisconsin, Milwaukee, WI, Unied

More information

MORPHOLOGICAL SEGMENTATION OF IMAGE SEQUENCES

MORPHOLOGICAL SEGMENTATION OF IMAGE SEQUENCES MORPHOLOGICAL SEGMENTATION OF IMAGE SEQUENCES B. MARCOTEGUI and F. MEYER Ecole des Mines de Paris, Cenre de Morphologie Mahémaique, 35, rue Sain-Honoré, F 77305 Fonainebleau Cedex, France Absrac. In image

More information

NEWTON S SECOND LAW OF MOTION

NEWTON S SECOND LAW OF MOTION Course and Secion Dae Names NEWTON S SECOND LAW OF MOTION The acceleraion of an objec is defined as he rae of change of elociy. If he elociy changes by an amoun in a ime, hen he aerage acceleraion during

More information

In Proceedings of CVPR '96. Structure and Motion of Curved 3D Objects from. using these methods [12].

In Proceedings of CVPR '96. Structure and Motion of Curved 3D Objects from. using these methods [12]. In Proceedings of CVPR '96 Srucure and Moion of Curved 3D Objecs from Monocular Silhouees B Vijayakumar David J Kriegman Dep of Elecrical Engineering Yale Universiy New Haven, CT 652-8267 Jean Ponce Compuer

More information

Shortest Path Algorithms. Lecture I: Shortest Path Algorithms. Example. Graphs and Matrices. Setting: Dr Kieran T. Herley.

Shortest Path Algorithms. Lecture I: Shortest Path Algorithms. Example. Graphs and Matrices. Setting: Dr Kieran T. Herley. Shores Pah Algorihms Background Seing: Lecure I: Shores Pah Algorihms Dr Kieran T. Herle Deparmen of Compuer Science Universi College Cork Ocober 201 direced graph, real edge weighs Le he lengh of a pah

More information

Parallax360: Stereoscopic 360 Scene Representation for Head-Motion Parallax

Parallax360: Stereoscopic 360 Scene Representation for Head-Motion Parallax Parallax360: Sereoscopic 360 Scene Represenaion for Head-Moion Parallax Bicheng Luo, Feng Xu, Chrisian Richard and Jun-Hai Yong (b) (c) (a) (d) (e) Fig. 1. We propose a novel image-based represenaion o

More information

STRING DESCRIPTIONS OF DATA FOR DISPLAY*

STRING DESCRIPTIONS OF DATA FOR DISPLAY* SLAC-PUB-383 January 1968 STRING DESCRIPTIONS OF DATA FOR DISPLAY* J. E. George and W. F. Miller Compuer Science Deparmen and Sanford Linear Acceleraor Cener Sanford Universiy Sanford, California Absrac

More information

Visual Perception as Bayesian Inference. David J Fleet. University of Toronto

Visual Perception as Bayesian Inference. David J Fleet. University of Toronto Visual Percepion as Bayesian Inference David J Flee Universiy of Torono Basic rules of probabiliy sum rule (for muually exclusive a ): produc rule (condiioning): independence (def n ): Bayes rule: marginalizaion:

More information

A Face Detection Method Based on Skin Color Model

A Face Detection Method Based on Skin Color Model A Face Deecion Mehod Based on Skin Color Model Dazhi Zhang Boying Wu Jiebao Sun Qinglei Liao Deparmen of Mahemaics Harbin Insiue of Technology Harbin China 150000 Zhang_dz@163.com mahwby@hi.edu.cn sunjiebao@om.com

More information

A High-Speed Adaptive Multi-Module Structured Light Scanner

A High-Speed Adaptive Multi-Module Structured Light Scanner A High-Speed Adapive Muli-Module Srucured Ligh Scanner Andreas Griesser 1 Luc Van Gool 1,2 1 Swiss Fed.Ins.of Techn.(ETH) 2 Kaholieke Univ. Leuven D-ITET/Compuer Vision Lab ESAT/VISICS Zürich, Swizerland

More information

MOTION DETECTORS GRAPH MATCHING LAB PRE-LAB QUESTIONS

MOTION DETECTORS GRAPH MATCHING LAB PRE-LAB QUESTIONS NME: TE: LOK: MOTION ETETORS GRPH MTHING L PRE-L QUESTIONS 1. Read he insrucions, and answer he following quesions. Make sure you resae he quesion so I don hae o read he quesion o undersand he answer..

More information

Evaluation and Improvement of Region-based Motion Segmentation

Evaluation and Improvement of Region-based Motion Segmentation Evaluaion and Improvemen of Region-based Moion Segmenaion Mark Ross Universiy Koblenz-Landau, Insiue of Compuaional Visualisics, Universiässraße 1, 56070 Koblenz, Germany Email: ross@uni-koblenz.de Absrac

More information

TrackNet: Simultaneous Detection and Tracking of Multiple Objects

TrackNet: Simultaneous Detection and Tracking of Multiple Objects TrackNe: Simulaneous Deecion and Tracking of Muliple Objecs Chenge Li New York Universiy cl2840@nyu.edu Gregory Dobler New York Universiy greg.dobler@nyu.edu Yilin Song New York Universiy ys1297@nyu.edu

More information

Effects needed for Realism. Ray Tracing. Ray Tracing: History. Outline. Foundations of Computer Graphics (Fall 2012)

Effects needed for Realism. Ray Tracing. Ray Tracing: History. Outline. Foundations of Computer Graphics (Fall 2012) Foundaions of ompuer Graphics (Fall 2012) S 184, Lecure 16: Ray Tracing hp://ins.eecs.berkeley.edu/~cs184 Effecs needed for Realism (Sof) Shadows Reflecions (Mirrors and Glossy) Transparency (Waer, Glass)

More information

AML710 CAD LECTURE 11 SPACE CURVES. Space Curves Intrinsic properties Synthetic curves

AML710 CAD LECTURE 11 SPACE CURVES. Space Curves Intrinsic properties Synthetic curves AML7 CAD LECTURE Space Curves Inrinsic properies Synheic curves A curve which may pass hrough any region of hreedimensional space, as conrased o a plane curve which mus lie on a single plane. Space curves

More information

Sam knows that his MP3 player has 40% of its battery life left and that the battery charges by an additional 12 percentage points every 15 minutes.

Sam knows that his MP3 player has 40% of its battery life left and that the battery charges by an additional 12 percentage points every 15 minutes. 8.F Baery Charging Task Sam wans o ake his MP3 player and his video game player on a car rip. An hour before hey plan o leave, he realized ha he forgo o charge he baeries las nigh. A ha poin, he plugged

More information

A Hierarchical Object Recognition System Based on Multi-scale Principal Curvature Regions

A Hierarchical Object Recognition System Based on Multi-scale Principal Curvature Regions A Hierarchical Objec Recogniion Sysem Based on Muli-scale Principal Curvaure Regions Wei Zhang, Hongli Deng, Thomas G Dieerich and Eric N Morensen School of Elecrical Engineering and Compuer Science Oregon

More information

Spline Curves. Color Interpolation. Normal Interpolation. Last Time? Today. glshademodel (GL_SMOOTH); Adjacency Data Structures. Mesh Simplification

Spline Curves. Color Interpolation. Normal Interpolation. Last Time? Today. glshademodel (GL_SMOOTH); Adjacency Data Structures. Mesh Simplification Las Time? Adjacency Daa Srucures Spline Curves Geomeric & opologic informaion Dynamic allocaion Efficiency of access Mesh Simplificaion edge collapse/verex spli geomorphs progressive ransmission view-dependen

More information

An Adaptive Spatial Depth Filter for 3D Rendering IP

An Adaptive Spatial Depth Filter for 3D Rendering IP JOURNAL OF SEMICONDUCTOR TECHNOLOGY AND SCIENCE, VOL.3, NO. 4, DECEMBER, 23 175 An Adapive Spaial Deph Filer for 3D Rendering IP Chang-Hyo Yu and Lee-Sup Kim Absrac In his paper, we presen a new mehod

More information

Landmarks: A New Model for Similarity-Based Pattern Querying in Time Series Databases

Landmarks: A New Model for Similarity-Based Pattern Querying in Time Series Databases Lmarks: A New Model for Similariy-Based Paern Querying in Time Series Daabases Chang-Shing Perng Haixun Wang Sylvia R. Zhang D. So Parker perng@cs.ucla.edu hxwang@cs.ucla.edu Sylvia Zhang@cle.com so@cs.ucla.edu

More information

A METHOD OF MODELING DEFORMATION OF AN OBJECT EMPLOYING SURROUNDING VIDEO CAMERAS

A METHOD OF MODELING DEFORMATION OF AN OBJECT EMPLOYING SURROUNDING VIDEO CAMERAS A METHOD OF MODELING DEFORMATION OF AN OBJECT EMLOYING SURROUNDING IDEO CAMERAS Joo Kooi TAN, Seiji ISHIKAWA Deparmen of Mechanical and Conrol Engineering Kushu Insiue of Technolog, Japan ehelan@is.cnl.kuech.ac.jp,

More information

Image Content Representation

Image Content Representation Image Conen Represenaion Represenaion for curves and shapes regions relaionships beween regions E.G.M. Perakis Image Represenaion & Recogniion 1 Reliable Represenaion Uniqueness: mus uniquely specify an

More information

4. Minimax and planning problems

4. Minimax and planning problems CS/ECE/ISyE 524 Inroducion o Opimizaion Spring 2017 18 4. Minima and planning problems ˆ Opimizing piecewise linear funcions ˆ Minima problems ˆ Eample: Chebyshev cener ˆ Muli-period planning problems

More information

Weighted Voting in 3D Random Forest Segmentation

Weighted Voting in 3D Random Forest Segmentation Weighed Voing in 3D Random Fores Segmenaion M. Yaqub,, P. Mahon 3, M. K. Javaid, C. Cooper, J. A. Noble NDORMS, Universiy of Oxford, IBME, Deparmen of Engineering Science, Universiy of Oxford, 3 MRC Epidemiology

More information

The Impact of Product Development on the Lifecycle of Defects

The Impact of Product Development on the Lifecycle of Defects The Impac of Produc Developmen on he Lifecycle of Rudolf Ramler Sofware Compeence Cener Hagenberg Sofware Park 21 A-4232 Hagenberg, Ausria +43 7236 3343 872 rudolf.ramler@scch.a ABSTRACT This paper invesigaes

More information

Algorithm for image reconstruction in multi-slice helical CT

Algorithm for image reconstruction in multi-slice helical CT Algorihm for image reconsrucion in muli-slice helical CT Kasuyuki Taguchi a) and Hiroshi Aradae Medical Engineering Laboraory, Toshiba Corporaion, 1385 Shimoishigami, Oawara, Tochigi 324-855, Japan Received

More information

Multi-Target Detection and Tracking from a Single Camera in Unmanned Aerial Vehicles (UAVs)

Multi-Target Detection and Tracking from a Single Camera in Unmanned Aerial Vehicles (UAVs) 2016 IEEE/RSJ Inernaional Conference on Inelligen Robos and Sysems (IROS) Daejeon Convenion Cener Ocober 9-14, 2016, Daejeon, Korea Muli-Targe Deecion and Tracking from a Single Camera in Unmanned Aerial

More information

Wheelchair-user Detection Combined with Parts-based Tracking

Wheelchair-user Detection Combined with Parts-based Tracking Wheelchair-user Deecion Combined wih Pars-based Tracking Ukyo Tanikawa 1, Yasuomo Kawanishi 1, Daisuke Deguchi 2,IchiroIde 1, Hiroshi Murase 1 and Ryo Kawai 3 1 Graduae School of Informaion Science, Nagoya

More information

Dynamic Depth Recovery from Multiple Synchronized Video Streams 1

Dynamic Depth Recovery from Multiple Synchronized Video Streams 1 Dynamic Deph Recoery from Muliple ynchronized Video reams Hai ao, Harpree. awhney, and Rakesh Kumar Deparmen of Compuer Engineering arnoff Corporaion Uniersiy of California a ana Cruz Washingon Road ana

More information

Real Time Integral-Based Structural Health Monitoring

Real Time Integral-Based Structural Health Monitoring Real Time Inegral-Based Srucural Healh Monioring The nd Inernaional Conference on Sensing Technology ICST 7 J. G. Chase, I. Singh-Leve, C. E. Hann, X. Chen Deparmen of Mechanical Engineering, Universiy

More information

Real-Time Avatar Animation Steered by Live Body Motion

Real-Time Avatar Animation Steered by Live Body Motion Real-Time Avaar Animaion Seered by Live Body Moion Oliver Schreer, Ralf Tanger, Peer Eiser, Peer Kauff, Bernhard Kaspar, and Roman Engler 3 Fraunhofer Insiue for Telecommunicaions/Heinrich-Herz-Insiu,

More information

DAGM 2011 Tutorial on Convex Optimization for Computer Vision

DAGM 2011 Tutorial on Convex Optimization for Computer Vision DAGM 2011 Tuorial on Convex Opimizaion for Compuer Vision Par 3: Convex Soluions for Sereo and Opical Flow Daniel Cremers Compuer Vision Group Technical Universiy of Munich Graz Universiy of Technology

More information

Image Based Computer-Aided Manufacturing Technology

Image Based Computer-Aided Manufacturing Technology Sensors & Transducers 03 by IFSA hp://www.sensorsporal.com Image Based Compuer-Aided Manufacuring Technology Zhanqi HU Xiaoqin ZHANG Jinze LI Wei LI College of Mechanical Engineering Yanshan Universiy

More information

Research Article Auto Coloring with Enhanced Character Registration

Research Article Auto Coloring with Enhanced Character Registration Compuer Games Technology Volume 2008, Aricle ID 35398, 7 pages doi:0.55/2008/35398 Research Aricle Auo Coloring wih Enhanced Characer Regisraion Jie Qiu, Hock Soon Seah, Feng Tian, Quan Chen, Zhongke Wu,

More information

SOT: Compact Representation for Triangle and Tetrahedral Meshes

SOT: Compact Representation for Triangle and Tetrahedral Meshes SOT: Compac Represenaion for Triangle and Terahedral Meshes Topraj Gurung and Jarek Rossignac School of Ineracive Compuing, College of Compuing, Georgia Insiue of Technology, Alana, GA ABSTRACT The Corner

More information

Detection and segmentation of moving objects in highly dynamic scenes

Detection and segmentation of moving objects in highly dynamic scenes Deecion and segmenaion of moving objecs in highly dynamic scenes Aurélie Bugeau Parick Pérez INRIA, Cenre Rennes - Breagne Alanique Universié de Rennes, Campus de Beaulieu, 35 042 Rennes Cedex, France

More information

Improving Occupancy Grid FastSLAM by Integrating Navigation Sensors

Improving Occupancy Grid FastSLAM by Integrating Navigation Sensors Improving Occupancy Grid FasSLAM by Inegraing Navigaion Sensors Chrisopher Weyers Sensors Direcorae Air Force Research Laboraory Wrigh-Paerson AFB, OH 45433 Gilber Peerson Deparmen of Elecrical and Compuer

More information

BI-TEMPORAL INDEXING

BI-TEMPORAL INDEXING BI-TEMPORAL INDEXING Mirella M. Moro Uniersidade Federal do Rio Grande do Sul Poro Alegre, RS, Brazil hp://www.inf.ufrgs.br/~mirella/ Vassilis J. Tsoras Uniersiy of California, Rierside Rierside, CA 92521,

More information

Rao-Blackwellized Particle Filtering for Probing-Based 6-DOF Localization in Robotic Assembly

Rao-Blackwellized Particle Filtering for Probing-Based 6-DOF Localization in Robotic Assembly MITSUBISHI ELECTRIC RESEARCH LABORATORIES hp://www.merl.com Rao-Blackwellized Paricle Filering for Probing-Based 6-DOF Localizaion in Roboic Assembly Yuichi Taguchi, Tim Marks, Haruhisa Okuda TR1-8 June

More information

AUTOMATIC 3D FACE REGISTRATION WITHOUT INITIALIZATION

AUTOMATIC 3D FACE REGISTRATION WITHOUT INITIALIZATION Chaper 3 AUTOMATIC 3D FACE REGISTRATION WITHOUT INITIALIZATION A. Koschan, V. R. Ayyagari, F. Boughorbel, and M. A. Abidi Imaging, Roboics, and Inelligen Sysems Laboraory, The Universiy of Tennessee, 334

More information

Graffiti Detection Using Two Views

Graffiti Detection Using Two Views Graffii Deecion Using wo Views Luigi Di Sefano Federico ombari Alessandro Lanza luigi.disefano@unibo.i federico.ombari@unibo.i alanza@arces.unibo.i Sefano Maoccia sefano.maoccia@unibo.i Sefano Moni sefano.moni3@sudio.unibo.i

More information

IntentSearch:Capturing User Intention for One-Click Internet Image Search

IntentSearch:Capturing User Intention for One-Click Internet Image Search JOURNAL OF L A T E X CLASS FILES, VOL. 6, NO. 1, JANUARY 2010 1 InenSearch:Capuring User Inenion for One-Click Inerne Image Search Xiaoou Tang, Fellow, IEEE, Ke Liu, Jingyu Cui, Suden Member, IEEE, Fang

More information

Robot localization under perceptual aliasing conditions based on laser reflectivity using particle filter

Robot localization under perceptual aliasing conditions based on laser reflectivity using particle filter Robo localizaion under percepual aliasing condiions based on laser refleciviy using paricle filer DongXiang Zhang, Ryo Kurazume, Yumi Iwashia, Tsuomu Hasegawa Absrac Global localizaion, which deermines

More information

Proceeding of the 6 th International Symposium on Artificial Intelligence and Robotics & Automation in Space: i-sairas 2001, Canadian Space Agency,

Proceeding of the 6 th International Symposium on Artificial Intelligence and Robotics & Automation in Space: i-sairas 2001, Canadian Space Agency, Proceeding of he 6 h Inernaional Symposium on Arificial Inelligence and Roboics & Auomaion in Space: i-sairas 00, Canadian Space Agency, S-Huber, Quebec, Canada, June 8-, 00. Muli-resoluion Mapping Using

More information

A Bayesian Approach to Video Object Segmentation via Merging 3D Watershed Volumes

A Bayesian Approach to Video Object Segmentation via Merging 3D Watershed Volumes A Bayesian Approach o Video Objec Segmenaion via Merging 3D Waershed Volumes Yu-Pao Tsai 1,3, Chih-Chuan Lai 1,2, Yi-Ping Hung 1,2, and Zen-Chung Shih 3 1 Insiue of Informaion Science, Academia Sinica,

More information

Computer representations of piecewise

Computer representations of piecewise Edior: Gabriel Taubin Inroducion o Geomeric Processing hrough Opimizaion Gabriel Taubin Brown Universiy Compuer represenaions o piecewise smooh suraces have become vial echnologies in areas ranging rom

More information

Mobile Robots Mapping

Mobile Robots Mapping Mobile Robos Mapping 1 Roboics is Easy conrol behavior percepion modelling domain model environmen model informaion exracion raw daa planning ask cogniion reasoning pah planning navigaion pah execuion

More information

Upper Body Tracking for Human-Machine Interaction with a Moving Camera

Upper Body Tracking for Human-Machine Interaction with a Moving Camera The 2009 IEEE/RSJ Inernaional Conference on Inelligen Robos and Sysems Ocober -5, 2009 S. Louis, USA Upper Body Tracking for Human-Machine Ineracion wih a Moving Camera Yi-Ru Chen, Cheng-Ming Huang, and

More information

MODEL BASED TECHNIQUE FOR VEHICLE TRACKING IN TRAFFIC VIDEO USING SPATIAL LOCAL FEATURES

MODEL BASED TECHNIQUE FOR VEHICLE TRACKING IN TRAFFIC VIDEO USING SPATIAL LOCAL FEATURES MODEL BASED TECHNIQUE FOR VEHICLE TRACKING IN TRAFFIC VIDEO USING SPATIAL LOCAL FEATURES Arun Kumar H. D. 1 and Prabhakar C. J. 2 1 Deparmen of Compuer Science, Kuvempu Universiy, Shimoga, India ABSTRACT

More information

Projection & Interaction

Projection & Interaction Projecion & Ineracion Algebra of projecion Canonical viewing volume rackball inerface ransform Hierarchies Preview of Assignmen #2 Lecure 8 Comp 236 Spring 25 Projecions Our lives are grealy simplified

More information

Simultaneous Localization and Mapping with Stereo Vision

Simultaneous Localization and Mapping with Stereo Vision Simulaneous Localizaion and Mapping wih Sereo Vision Mahew N. Dailey Compuer Science and Informaion Managemen Asian Insiue of Technology Pahumhani, Thailand Email: mdailey@ai.ac.h Manukid Parnichkun Mecharonics

More information

Design Alternatives for a Thin Lens Spatial Integrator Array

Design Alternatives for a Thin Lens Spatial Integrator Array Egyp. J. Solids, Vol. (7), No. (), (004) 75 Design Alernaives for a Thin Lens Spaial Inegraor Array Hala Kamal *, Daniel V azquez and Javier Alda and E. Bernabeu Opics Deparmen. Universiy Compluense of

More information

Curves & Surfaces. Last Time? Today. Readings for Today (pick one) Limitations of Polygonal Meshes. Today. Adjacency Data Structures

Curves & Surfaces. Last Time? Today. Readings for Today (pick one) Limitations of Polygonal Meshes. Today. Adjacency Data Structures Las Time? Adjacency Daa Srucures Geomeric & opologic informaion Dynamic allocaion Efficiency of access Curves & Surfaces Mesh Simplificaion edge collapse/verex spli geomorphs progressive ransmission view-dependen

More information

Coded Caching with Multiple File Requests

Coded Caching with Multiple File Requests Coded Caching wih Muliple File Requess Yi-Peng Wei Sennur Ulukus Deparmen of Elecrical and Compuer Engineering Universiy of Maryland College Park, MD 20742 ypwei@umd.edu ulukus@umd.edu Absrac We sudy a

More information

MOTION TRACKING is a fundamental capability that

MOTION TRACKING is a fundamental capability that TECHNICAL REPORT CRES-05-008, CENTER FOR ROBOTICS AND EMBEDDED SYSTEMS, UNIVERSITY OF SOUTHERN CALIFORNIA 1 Real-ime Moion Tracking from a Mobile Robo Boyoon Jung, Suden Member, IEEE, Gaurav S. Sukhame,

More information

Last Time: Curves & Surfaces. Today. Questions? Limitations of Polygonal Meshes. Can We Disguise the Facets?

Last Time: Curves & Surfaces. Today. Questions? Limitations of Polygonal Meshes. Can We Disguise the Facets? Las Time: Curves & Surfaces Expeced value and variance Mone-Carlo in graphics Imporance sampling Sraified sampling Pah Tracing Irradiance Cache Phoon Mapping Quesions? Today Moivaion Limiaions of Polygonal

More information

Real time 3D face and facial feature tracking

Real time 3D face and facial feature tracking J Real-Time Image Proc (2007) 2:35 44 DOI 10.1007/s11554-007-0032-2 ORIGINAL RESEARCH PAPER Real ime 3D face and facial feaure racking Fadi Dornaika Æ Javier Orozco Received: 23 November 2006 / Acceped:

More information

Robust Multi-view Face Detection Using Error Correcting Output Codes

Robust Multi-view Face Detection Using Error Correcting Output Codes Robus Muli-view Face Deecion Using Error Correcing Oupu Codes Hongming Zhang,2, Wen GaoP P, Xilin Chen 2, Shiguang Shan 2, and Debin Zhao Deparmen of Compuer Science and Engineering, Harbin Insiue of Technolog

More information

LOW-VELOCITY IMPACT LOCALIZATION OF THE COMPOSITE TUBE USING A NORMALIZED CROSS-CORRELATION METHOD

LOW-VELOCITY IMPACT LOCALIZATION OF THE COMPOSITE TUBE USING A NORMALIZED CROSS-CORRELATION METHOD 21 s Inernaional Conference on Composie Maerials Xi an, 20-25 h Augus 2017 LOW-VELOCITY IMPACT LOCALIZATION OF THE COMPOSITE TUBE USING A NORMALIZED CROSS-CORRELATION METHOD Hyunseok Kwon 1, Yurim Park

More information

Network management and QoS provisioning - QoS in Frame Relay. . packet switching with virtual circuit service (virtual circuits are bidirectional);

Network management and QoS provisioning - QoS in Frame Relay. . packet switching with virtual circuit service (virtual circuits are bidirectional); QoS in Frame Relay Frame relay characerisics are:. packe swiching wih virual circui service (virual circuis are bidirecional);. labels are called DLCI (Daa Link Connecion Idenifier);. for connecion is

More information

Vision-Based Traffic Measurement System

Vision-Based Traffic Measurement System *J. M. Wang, **S. L. Chang, **Y. C. Chung, and **S. W. Chen Deparmen of Informaion and Compuer Educaion *Naional Taiwan Universiy **Naional Taiwan Normal Universiy Taipei, Taiwan schen@csie.nnu.edu.w Absrac

More information

PART 1 REFERENCE INFORMATION CONTROL DATA 6400 SYSTEMS CENTRAL PROCESSOR MONITOR

PART 1 REFERENCE INFORMATION CONTROL DATA 6400 SYSTEMS CENTRAL PROCESSOR MONITOR . ~ PART 1 c 0 \,).,,.,, REFERENCE NFORMATON CONTROL DATA 6400 SYSTEMS CENTRAL PROCESSOR MONTOR n CONTROL DATA 6400 Compuer Sysems, sysem funcions are normally handled by he Monior locaed in a Peripheral

More information

MATH Differential Equations September 15, 2008 Project 1, Fall 2008 Due: September 24, 2008

MATH Differential Equations September 15, 2008 Project 1, Fall 2008 Due: September 24, 2008 MATH 5 - Differenial Equaions Sepember 15, 8 Projec 1, Fall 8 Due: Sepember 4, 8 Lab 1.3 - Logisics Populaion Models wih Harvesing For his projec we consider lab 1.3 of Differenial Equaions pages 146 o

More information

Section 2. Mirrors and Prism Systems

Section 2. Mirrors and Prism Systems Secion 2 Mirrors and Prism Sysems 2-1 Plane Mirrors Plane mirrors are used o: Produce a deviaion Fold he opical pah Change he image pariy Each ray from he objec poin obeys he law of reflecion a he mirror

More information

Video-Based Face Recognition Using Probabilistic Appearance Manifolds

Video-Based Face Recognition Using Probabilistic Appearance Manifolds Video-Based Face Recogniion Using Probabilisic Appearance Manifolds Kuang-Chih Lee Jeffrey Ho Ming-Hsuan Yang David Kriegman klee10@uiuc.edu jho@cs.ucsd.edu myang@honda-ri.com kriegman@cs.ucsd.edu Compuer

More information

Simultaneous Precise Solutions to the Visibility Problem of Sculptured Models

Simultaneous Precise Solutions to the Visibility Problem of Sculptured Models Simulaneous Precise Soluions o he Visibiliy Problem of Sculpured Models Joon-Kyung Seong 1, Gershon Elber 2, and Elaine Cohen 1 1 Universiy of Uah, Sal Lake Ciy, UT84112, USA, seong@cs.uah.edu, cohen@cs.uah.edu

More information

COSC 3213: Computer Networks I Chapter 6 Handout # 7

COSC 3213: Computer Networks I Chapter 6 Handout # 7 COSC 3213: Compuer Neworks I Chaper 6 Handou # 7 Insrucor: Dr. Marvin Mandelbaum Deparmen of Compuer Science York Universiy F05 Secion A Medium Access Conrol (MAC) Topics: 1. Muliple Access Communicaions:

More information

Motion Level-of-Detail: A Simplification Method on Crowd Scene

Motion Level-of-Detail: A Simplification Method on Crowd Scene Moion Level-of-Deail: A Simplificaion Mehod on Crowd Scene Absrac Junghyun Ahn VR lab, EECS, KAIST ChocChoggi@vr.kais.ac.kr hp://vr.kais.ac.kr/~zhaoyue Recen echnological improvemen in characer animaion

More information

A Fast Non-Uniform Knots Placement Method for B-Spline Fitting

A Fast Non-Uniform Knots Placement Method for B-Spline Fitting 2015 IEEE Inernaional Conference on Advanced Inelligen Mecharonics (AIM) July 7-11, 2015. Busan, Korea A Fas Non-Uniform Knos Placemen Mehod for B-Spline Fiing T. Tjahjowidodo, VT. Dung, and ML. Han Absrac

More information

3-D Object Modeling and Recognition for Telerobotic Manipulation

3-D Object Modeling and Recognition for Telerobotic Manipulation Research Showcase @ CMU Roboics Insiue School of Compuer Science 1995 3-D Objec Modeling and Recogniion for Teleroboic Manipulaion Andrew Johnson Parick Leger Regis Hoffman Marial Heber James Osborn Follow

More information

Analysis of Various Types of Bugs in the Object Oriented Java Script Language Coding

Analysis of Various Types of Bugs in the Object Oriented Java Script Language Coding Indian Journal of Science and Technology, Vol 8(21), DOI: 10.17485/ijs/2015/v8i21/69958, Sepember 2015 ISSN (Prin) : 0974-6846 ISSN (Online) : 0974-5645 Analysis of Various Types of Bugs in he Objec Oriened

More information

Optimal Crane Scheduling

Optimal Crane Scheduling Opimal Crane Scheduling Samid Hoda, John Hooker Laife Genc Kaya, Ben Peerson Carnegie Mellon Universiy Iiro Harjunkoski ABB Corporae Research EWO - 13 November 2007 1/16 Problem Track-mouned cranes move

More information

Track-based and object-based occlusion for people tracking refinement in indoor surveillance

Track-based and object-based occlusion for people tracking refinement in indoor surveillance Trac-based and objec-based occlusion for people racing refinemen in indoor surveillance R. Cucchiara, C. Grana, G. Tardini Diparimeno di Ingegneria Informaica - Universiy of Modena and Reggio Emilia Via

More information

A GRAPHICS PROCESSING UNIT IMPLEMENTATION OF THE PARTICLE FILTER

A GRAPHICS PROCESSING UNIT IMPLEMENTATION OF THE PARTICLE FILTER A GRAPHICS PROCESSING UNIT IMPLEMENTATION OF THE PARTICLE FILTER ABSTRACT Modern graphics cards for compuers, and especially heir graphics processing unis (GPUs), are designed for fas rendering of graphics.

More information

Quantitative macro models feature an infinite number of periods A more realistic (?) view of time

Quantitative macro models feature an infinite number of periods A more realistic (?) view of time INFINIE-HORIZON CONSUMPION-SAVINGS MODEL SEPEMBER, Inroducion BASICS Quaniaive macro models feaure an infinie number of periods A more realisic (?) view of ime Infinie number of periods A meaphor for many

More information

Representing Non-Manifold Shapes in Arbitrary Dimensions

Representing Non-Manifold Shapes in Arbitrary Dimensions Represening Non-Manifold Shapes in Arbirary Dimensions Leila De Floriani,2 and Annie Hui 2 DISI, Universiy of Genova, Via Dodecaneso, 35-646 Genova (Ialy). 2 Deparmen of Compuer Science, Universiy of Maryland,

More information

Real-time 2D Video/3D LiDAR Registration

Real-time 2D Video/3D LiDAR Registration Real-ime 2D Video/3D LiDAR Regisraion C. Bodenseiner Fraunhofer IOSB chrisoph.bodenseiner@iosb.fraunhofer.de M. Arens Fraunhofer IOSB michael.arens@iosb.fraunhofer.de Absrac Progress in LiDAR scanning

More information

Real-Time Non-Rigid Multi-Frame Depth Video Super-Resolution

Real-Time Non-Rigid Multi-Frame Depth Video Super-Resolution Real-Time Non-Rigid Muli-Frame Deph Video Super-Resoluion Kassem Al Ismaeil 1, Djamila Aouada 1, Thomas Solignac 2, Bruno Mirbach 2, Björn Oersen 1 1 Inerdisciplinary Cenre for Securiy, Reliabiliy, and

More information

Detection Tracking and Recognition of Human Poses for a Real Time Spatial Game

Detection Tracking and Recognition of Human Poses for a Real Time Spatial Game Deecion Tracking and Recogniion of Human Poses for a Real Time Spaial Game Feifei Huo, Emile A. Hendriks, A.H.J. Oomes Delf Universiy of Technology The Neherlands f.huo@udelf.nl Pascal van Beek, Remco

More information

ACQUIRING high-quality and well-defined depth data. Online Temporally Consistent Indoor Depth Video Enhancement via Static Structure

ACQUIRING high-quality and well-defined depth data. Online Temporally Consistent Indoor Depth Video Enhancement via Static Structure SUBMITTED TO TRANSACTION ON IMAGE PROCESSING 1 Online Temporally Consisen Indoor Deph Video Enhancemen via Saic Srucure Lu Sheng, Suden Member, IEEE, King Ngi Ngan, Fellow, IEEE, Chern-Loon Lim and Songnan

More information

J. Vis. Commun. Image R.

J. Vis. Commun. Image R. J. Vis. Commun. Image R. 20 (2009) 9 27 Conens liss available a ScienceDirec J. Vis. Commun. Image R. journal homepage: www.elsevier.com/locae/jvci Face deecion and racking using a Boosed Adapive Paricle

More information

We are IntechOpen, the world s leading publisher of Open Access books Built by scientists, for scientists. International authors and editors

We are IntechOpen, the world s leading publisher of Open Access books Built by scientists, for scientists. International authors and editors We are InechOpen, he world s leading publisher of Open Access books Buil by scieniss, for scieniss 4,000 116,000 120M Open access books available Inernaional auhors and ediors Downloads Our auhors are

More information

An Iterative Scheme for Motion-Based Scene Segmentation

An Iterative Scheme for Motion-Based Scene Segmentation An Ieraive Scheme for Moion-Based Scene Segmenaion Alexander Bachmann and Hildegard Kuehne Deparmen for Measuremen and Conrol Insiue for Anhropomaics Universiy of Karlsruhe (H), 76 131 Karlsruhe, Germany

More information

High Resolution Passive Facial Performance Capture

High Resolution Passive Facial Performance Capture High Resoluion Passive Facial Performance Capure Derek Bradley1 Wolfgang Heidrich1 Tiberiu Popa1,2 Alla Sheffer1 1) Universiy of Briish Columbia 2) ETH Zu rich Figure 1: High resoluion passive facial performance

More information

A MRF formulation for coded structured light

A MRF formulation for coded structured light A MRF formulaion for coded srucured ligh Jean-Philippe Tardif Sébasien Roy Déparemen d informaique e recherche opéraionnelle Universié de Monréal, Canada {ardifj, roys}@iro.umonreal.ca Absrac Mulimedia

More information