[PDF] Towards Automatic 3D Reconstruction from 2D Floorplan Image PDF flat.pdf

Reconstruction of 3D model representation from a 2D image(s) is proven to be a diffi- cult task In this paper, we present a simple solution to the restricted case of

Python Photogrammetry Toolbox (PPT) [Ref S4] - a project to integrate Bundler, CMVS and PMVS into an open-source photogrammetry toolbox by the

[PDF] Evaluation of Dense 3D Reconstruction from 2D Face Images in the

3D face reconstruction from a single 2D image in the wild To this end, we In addition to the dataset, we provide a standard protocol as well as a Python

[PDF] Python Photogrammetry Toolbox - HAL-ENPC

18 jui 2013 · Calibration is performed with Bundler while dense reconstruction is done camera center to 3D points and the 2d image plane coordinates

[PDF] Methods for 3D Reconstruction from Multiple Images

The image sequence is available on Long Quan's webpage: 2D pixel 3D ray [Lhuillier 02] ECCV'02, Quasi-Dense Reconstruction from Image Sequence

[PDF] Towards Automatic 3D Reconstruction from 2D Floorplan Image

Reconstruction of 3D model representation from a 2D image(s) is proven to be a diffi- cult task In this paper, we present a simple solution to the restricted case of

[PDF] 3D Reconstruction and Camera Calibration from 2D Images

A 3D reconstruction technique from stereo images is presented that needs minimal intervention 2D —Two-dimensional space, eg: the image plane 3D

Department of Computer Science ? EngineeringThe Chinese University of Hong Kong? Shatin? N?T?? Hong Kongfshor?cse?cuhk?edu?hkgAbstract

Reconstruction of 3D mo del representationfrom a 2D image?s? is proven to b e a di??cult task?In this pap er? we presentasimplesolutionto therestrictedcaseofgeneratingmo del description of a building solely from its2D ?o orplan?The motivation of this researchis that i? a lot of existing buildings has only

the ?o orplans remained? ii? In general a lot ofe?orts i?e human lab or work? haveto be made inorder to build the 3D mo del? We prop osed twoapproaches with di?erent degree of automation:The ?rst approachoperatesby an user analyzesand breaks down the ?o orplan into a number ofbuilding primitiv

es? The co ordinate informationare thenmanuallyrecordedandinputto aprogram whichwillgenerates the 3D mo deldescriptions i?e?a p opular 3D mo del ?le suchas3D Studioetc?Another approach appliesimage pro cessing techniques in which the userselects and changes the lines in the ?o orplan intodi?erent color co ded areas?These areas againrepresent di?erent 3D primitives similar to the?rst approach? A raster to vector program then

converts these primitives into 3D format?Thelatter approachinvolves only minimal humane?ort?Using our approaches? an ordinary usercan pro duce a 3D mo del of a building in a matter

of seconds? We present some of the results anddiscuss the relative merits of the two approaches?The presented algorithms are well suited for datageneration in architectural walkthrough on oldbuildings and 3D games development?keyword:

3D modeling? block?basedapproach?raster to vector conversion? color?codedblocks

1Intro ductionGiven the rapid advances in mo deling technolo?gies in recentyears? it is relatively easy for an av?erage user to build a 3D mo del of a building using

any p opular mo deling to ols suchasMayaor 3DStudio?However for buildings that exist yearsago? what we got are only their ?o orplans? whichis in general a 2D representation of the buildingunder varying assumptions? A lot of such build?ings are of historic as well as educational imp or?tance if one can navigate and examine its struc?ture ?1?? To ?recover? such buildings in the virtualspace? tedious mo deling pro cedures are usuallyrequired?A ?o orplan has the feature of abstracting a lot

of spacial information in a 2D p ostulation?In

general a ?o orplan consists of a numb er of linesof various widths p ostulating the top view of abuilding examined? The lines designate walls ofdi?erent thickness? In actual buildings? esp eciallythose used for residential purp ose? it happ ens thatop enings on the walls? i?e? windows? are needed?In order to cater for the ab ove situation? a sp ecialsignature? say a pair of thin lines is used to denotethat an op ening should b e placed there? Dep endson the building nature? usually a set of rules orsignatures is used to represent the deviation fromthe normal simple wall cases?With the aid of current mo delling package? onehave to trace the lines on the ?o orplan? whichrepresents the walls and instruct the package tobuild the corresp onding 3D representations?Atypical way is to use a graphic?tablet?and?prob eapproachtomanually trace the lines on the draw?1

Figure 1: A typical ?o orplan of a buildinging so as to enter the co ordinate information intothe computer? During this pro cess? the mo dellerhas to take care of the various asp ects of the planby noting attributes? such as height? thickness?of the lines and tell the package such informa?tion? The result is that much time is needed forconverting a ?o orplan into its 3D representation?Not to mention the more additional e?orts to fur?ther re?ne the 3D mo dels such as texturing andlighting?Another disadvantage is that a skilledp ersonnel is required in order to convert the em?b edded spacial signature into corresp onding 3Dinformation?2Our ApproachIn examining the layout of a ?o orplan? a linesegment on a ?o orplan denotes a blo ckorientedvertically in the 3D space? A simple approachistherefore to register the co ordinates of the four

points de?ning each line on a ?o orplan and gen?erate the 3D representation using a normal in y?axis? By recursively pro cessing all the lines in a?o orplan? we should b e able to generate the 3Ddescription of the whole building?This approach su?ers from the following draw?backs? Firstly there are various line signatures in?side a typical ?o orplan which designate di?erentconstructs in 3D space? for example a windowedblo ck? Decomp osing a ?o orplan solely into ver?tical blo cks cannot faithfully repro duce the cor?rect 3D representation? In addition? eachverti?cal blo ck generated typically will have faces thatnever b e seen? for example? the base of a verticalblo ckaswell as part of the faces in a L?shap edlines which are b eing o ccluded by its orthogonallyaligned neighb or?These faces will a?ect b oththe memory required during program executionand the rendering p erformance? Finally these re?

dundant faces also create visual artifacts on mostmo dern rendering engines?

2?1Blo ck Based ApproachIt is inevitable that the vertical blo ckbasedapproachmentioned in previous section? in spiteof the numb er of drawbacks? is feasible and easyto start with?We extend the idea of buildingblo cks and come up with an approach whichhasthe simplicity advantage but avoiding the draw?backs mentioned?By examining extensively the di?erent line sig?natures in a ?o orplan as in ?g?1? wederive sev?eral basic building blo cks encountered in a usualbuilding? In our example? the target is the interiorof a ?at? We emphasize here what we presentedis a framework and in practice more basic build?ing blo cks can b e designed given the appropriateapplication arena? Moreover webelieved that theblo cks we are going to present include most of theelements required in a ?o orplan?In the following discussions? we assumed thata ?o orplan denotes the top view of the building?which is thex?zplane and the vertical axis is they?axis??The simplest elementary blo ck is the verticalblo ck? we termed it ?Box??ABoxcan b efully describ ed by 6 parameters??g? 2?:2

Figure 2: Various building blo cksThe p osition of the reference p oint?3??The dimensions? or size of the b oxinthreeaxis?3???To cater for b oxes rotated at an angle ab outthey?axis? we add a generalized b ox notionand termed it ?G?Box??AG?Boxis fullyrepresented by 10 parameters ??g? 2?:X?? ? Z?co ordinates for the four vertices?8??lower and upp er Y?co ordinates??Windows? or related structures? are p opu?lar in man?made building?To representawindow? weintro duce the ?donut? blo ck? Adonutis a blo ck with an op ening in its inte?rior and is mo deled by 10 parameters??g? 2?:The outer lower left co ordinates?3? of thewhole blo ck?The dimensions of the whole blo ck?3??The co ordinates of the lower left corner ofinner blo ck?2??The dimensions of inner blo ck?2?fSee ?g?2g??Do or structures are also frequently seen in a?o orplan? Weintro duce an ?Ann?block ?See

?g?2? to mo del this?Annblo ck has fewerparameters ?8? thandonut:The lower left co ordinates?3? of the wholeblo ck?The dimensions of the whole blo ck?3??The dimensions of inner blo ck?2???A simple extension ofAnnis the ?You?blo ck?whichisbasicallyaninvertedAnn?See ?g?2?:?AcombinationofBoxwhichisp opu?lar in ?o orplan construct is theL?shap edblo ck ??g?2?? AnL?blo ck is basically a sub?traction of a smaller b ox from a larger one?We use 8 parameters for this blo ck:The lower left co ordinates?3? of the wholeblo ck?The dimensions of the whole blo ck?3??The p osition of inner corner of L?2???As inG?Box?wehave the generalized ver?sions of all the ab ove blo cks which need moreparameters to sp ecify? As these are trivial ex?tensions? we will not detail their implemen?tations here?The op erations in utilizing the ab ove buildingblo cks to parameterize a ?o orplan involves treat?ing the line complex of the whole ?o orplan as a

combinations of the blo cktyp es and decomp osingthem and recording the parameters? The detailalgorithm works as follow:Block BasedAlgorithm for building the 3D rep?resentation from a ?oorplan1? Overlay a uniform grid on the ?o orplan im?age? The spacing of the grid can convenientlybe chosen the same as the scale noted on theplan? In this case? the recovered 3D represen?tation will thus have exactly the same scaleas in human world?2? Select a p oint on the grid as the origin? Man?ually record the co ordinates of vertices of alllines in the plan?3

Figure 3: Breakdown of a p ortion of ?o orplan intoblo cks3? Based on the di?erent line signatures in theplan? map a suitable blo cktyp e to the lines?4? All the blo ck information are then enteredinto a program which will generate the cor?resp onding 3D data ?le?5? The user can then p erform an interactivewalkthrough on the resulting data ?le so asto re?ne the mo del? This iterative re?nementpro cess can b e p erformed on the data ?legenerated by the program in step 4? Addi?tonal touch can also b e done byvarious moresophisticated package suchas3D StudioorMaya?Atypical result of applying the algorithm toa ?o orplan is shown in ?g?3? In this ?gure? aL?blo ck is ?rst setted for the upp er left lines? Sub?jected to this choice? the lower left blo ckcanbechosen as anotherLor aBox? But since the lineon the lower righthasadooronit?whichcalledfor anAnnthere? thus a singleBoxis selected?For the upp er right p ortion? here we found a lineon which there should b e an op ening on it ?theair conditioner ?s p osition?? thus adonutis cho?sen? As can b e seen in this example? the setting

of blo cks? choices can b e varied during the appli?cation of the algorithm and there is no unique so?lution to it? The b est result can only b e achievedthrough exp eriments and the understanding ofthe relative merits of the blo cks?2?2Color?co ded Polylines ApproachTheblo ck?basedapproachprovidesafastmetho d to recover the 3D representation from a?o orplan image? The drawback is that it still re?quires a lab or intensive stage of collecting co ordi?nates of the lines in the plan? To further reducethe time needed? we devise a more automated ap?proach? In this approach? the di?erent signaturesin a ?o orplan are co ded as various unique colors?The algorithm is shown as b elow?Color?coded Polylines algorithm:Convert a?oorplan image into 3D representation1? The ?o orplan image? which is assumed to b ein gray scale? is prop erly thresholded into abinary image so that only the lines represent?ing the construct are left? It is also b etter toremoveany textual patterns reside?2? Di?erent structural signatures in the result?ing image are colored according to theirco des? for example? a line segment whichhasan op ening is selected and the whole segmentis colored into blue?3? A program read the colored image? and based

on the di?erent color? ?parse? the image intoseveral binary versions of the image? Eachofthese images contains only the p ortion whichb elongs to that particular construct? for ex?ample a windowed one?4? Another program reads these images in turn?

and converts the binary image into a vec?tored one? i?e? a p olyline description of theoutlines of the colored region?5? A ?nal program is then used to generate the

3D description i?e?the 3D formatted ?le?from the p olyline ?les generated in previous

stage? The program will take care of di?erentconstructs and their corresp onding descrip?tions in 3D space?4

Figure 4: FlatEd ? the program to build a blo ck?based 3D ?leAs one can readily observe? minimal human ef?fort ?stage 1? is involved in the ab ove algorithm?The user only needed to i? prop erly control thethresholding pro cess to minimize the unneededp olylines from noise in image? ii? Color the imageaccording to the signature of di?erent constructs?The color?co ded algorithm has various advan?tages over the blo ck?based one?The most im?p ortant one is that the user now no longer hasto go through the tedious stage of measurementof vertex co ordinates? This results in signi?cantsp eedup of the coversion pro cess? In addition? theusage of color co de guaranteed more ?exibilityinde?ning one?s ?o orplan requirement? Finally thecolor?co ded algorithm do es not require the userto b e pro?cient in di?erent ?o orplan constructsi?e? the breakdown pro cess?A little drawback of the color?co ded algorithmis that the parameters of di?erent constructs haveto b e ?xed prior to pro cessing? thus limiting thefreedom of the user to mo dify the outlo ok of thegenerated 3D ?les?However? this problem caneasily b e solved by using more sophisticated mo d?eling package suchasAutoCadorMaya?3Implementation and DiscussionWe implemented b oth approaches and testedthe generated 3D ?les onCrystal Space? whichisa free 3D engine written by Jorrit ?6? and oth?ers?The engine allows interactivewalkthrough

Figure 5: FlatEd allows the user to ne tune thegenerated mo dels.of the generated le so that one can immediately

insp ect the mo del generated. In principle, giventhat the p olygonal description of most 3D le for-

mats are basically the same, it is relatively easyfor one to mo dify the program to generate le inother 3D formats. We brie

y describ e b elowourimplementation and discuss p ossible extensions.3?1Blo ck?based ApproachThe parameters of every blo ck app eared inthe ?o orplan is recorded and entered into a pro?gram ??g? 4? whichwe called ?FlatEd?? This pro?gram is written inVisual Basicdue to its supp ortin various data entry widgets? The program alsoallows the user to ?ne tune the blo cks such as tex?turing parameters ??g?5??FlatEdinterfaces witha COM server which is written inPython?5?? ThePythonserver will generate the needed world ?lei?e?3D representation inCrystal Spaceformat?A screen shot of one of the generated scenes isshown in ?g? 7?3?2Color?co ded Polylines ApproachThe output after thresholding and colorization

is shown in ?g? 8? In this ?gure? a black line repre?sents an ordinary wall? green blo cks corresp ond tobay windows? a yellow blo ck denotes the p ositionof light sources? whereas blue one is the windowedp ortion? The fact that light sources are explicitlyrepresented is solely an advantage arised from the5

Figure 6: A typical ?o orplan converted into a col?lection of blo cks? Figure 7: View of g. 6 with texture and lighting.use of color co de.

During stage 2, weneedtoconvert the binaryimage into a vectored format, i.e. the p olylines.We use Ras2Vec (Raster to vector conversion)program which is written byDavide [4].Thep olylines generated is shown in g. 9.The conversion of dierent p olylines into a sin-gle 3D le is p erformed by a program inPython

Figure 8: A ?o orplan in color co ded format?scripts? A p oint to note is that the p olygons gen?erated are in general required to b e convex form?which is a basic requirement of most 3D render?ing engines? However it is clear that sometimesthe p olygons generated from our approacharenot necessarily convex?As noted in ?g? 8? thebay window area on upp er right is a concave one?We use the p olygon decomp osition algorithm by?3??The algorithm will break down a complex?concave p olygons into a numb er of convex p oly?gons? We adopt the program written by Ofer ?2?into aPythonversion whichisintegrated into ourPython COMserver mentioned in blo ck?based ap?proach? The resulting p olygons are added to the3D representation?An application of the color?co ded approachto?o orplans representing larger area is shown in?g? 10? Fig? 10 is a rendered view of a develop?ment pro ject in whichseveral buildings as well asthe roads and trees are placed using the di?erentcolors in the plan?6

Figure 9:Flo orplan converted to p olyline byras2Vec? Figure 10: Another application of color-co ded ap-proach to

o orplan.3?3DiscussionIn this pap er? we presenttwo di?erentap?proaches to generating 3D data from a ?o orplanimage? Both approaches aimed at quickly gener?ating a 3D mo del from a ?o orplan image? whichisa ?rst step in interactive 3D applications suchasgames and architectural walkthrough for ancientbuildings?Usually for such applications? b eside

lowering the cost of mo del data generation? onewould also wanttomodify as well as ?ne tune theresults so that more eye?candies can b e providedfor ?nal applications?In our exp eriments? b oth approaches can gen?erate 3D mo dels whichhave high precision withresp ect to original ?o orplan? However from ourexp eriences? the two approaches eachhave theirown relative merits and drawbacks as well?The blo ck?based approach has the drawbackoftedious lab or in vertices information preparation?In addition? it requires the user to have quite in?depth knowledge on di?erentblocktyp es so as tobuild a mo del of seamless joints b etween blo cks?On the other hand? blo ck based approach allowsits user to p erform more ?ne?tuning op erationsto the resulting mo dels such as adding furnitures?or even adding sp ecial features on a particularp olygon??g? 11??An even more attractivefea?ture is that it allow the user to easily diveintothe generated data ?le to ?ne tune the mo del forsp ecial e?ect such as p erformance tuning by re?moving unused p olygons? A typical ?o orplan asin ?g? 8? which results in 637 p olygons requiresapproximately 2 hours to generate the 3D mo del?le?For the color?co ded approach? relatively littleknowledge is required for a user to generate a3D mo del? This op ens up more avenues for thisapproach?One interesting application is an in?tegrated environment where the user just has todraw with di?erent colored p ens and the programwill build a mo del for interactive tour? The appli?cation should b e well suited as a creativitytoolforchildren? etc? In addition? the time used to builda 3D mo del is signi?cantly reduced? for example?the 3D mo del is generated in half an hour? includ?ing the image pro cessing time and color co ding?The drawbacks of the color?co ded approacharethat lesser ?ne?tune op erations as well as ?exibil?ity in mo deling is allowed? In addition? the gen?erated ?le is di?cult to edit? However wenotedthat the user can still edit the resulting data ?le

in more sophisticated mo deling packages?In view of the goal of minimizing the time in construction of mo del from a ?o orplan? wecon?7

Figure 11:Blo ck based approach allows easymo di?cations as well as analytical placementofother items?cluded that we are successful in signi?cantly re?ducing the time needed? As a simple b enchmarkof the p erformance of our algorithm? we notedthat the ?o orplan presented in ?g? 1 originallyneed a time of approximately two to three work?ing days to build the resulting mo del using plainestimation and hand editing of data ?le? in con?trast with the time rep orted ab ove?One mayargue that using sophisticated mo deling package?one can signi?cantly reduce the mo deling time?But as the mo deling package nowaday are writ?ten with generic purp oses in mind? it would needquite extensivework in order to build the samemo del? Our algorithm takes advantages of the or?thogonal feature of ?o orplan? and the redundantconstructs in man?made building? The resultingalgorithm can thus save more times in comparisonto more general mo deling approaches?4ConclusionTwo e?cient approaches to construct a 3Dmo del from a ?o orplan image are discussed? The?rst approachoperates by ?tting a set of build?ing blo cks to the ?o orplan image?The secondapproach used di?erent colors to representvari?ous constructs inside a ?o orplan image and then

convert the raster image into a numb er of p oly?lines? Each p olyline is then used to generate a setof p olygons which represent that particular blo ck?

Figure 12: Some screen shots of the buildings cre-ated.Relative merits and drawbacks of each algorithmare discussed and various results are presented.5AcknowledgementThe author would like to thank Francis Lee forsuggesting the idea of blo cks based approach andvaluable helps throughout the pro ject?8

References?1? Historical ?o orplan emp orium?http:??www?b?ware?com?hive?fplans?index?htm??2? OferBelinsky?Minimalcon?vexp olygondecomp ositiondemo?http:??www?math?tau?ac?il? sariel?TA?wcg98b??convex

decomp/Default.html??3? J?M? Keil? Decomp osing a p olygon into sim?pler comp onents?SIAM JC? 14:799?817? 1985??4? Davide Lib enzi?Raster to vector conversionprogram?http:??www?maticad?it?davide???5? AndyRobinson?Emb eddingpythoninvisual basic ? delphi ? p owerbuilder apps?http:??www?robanal?demon?co?uk?demos?pyvb?index?html??6? Jorrit Tyb erghein?Crystal space: a free 3dengine?http:??crystal?linuxgames?com??9

quotesdbs_dbs17.pdfusesText_23

[PDF] [PDF] Towards Automatic 3D Reconstruction from 2D Floorplan Image

[PDF] 3D Reconstruction from Multiple Images - Stanford Computational