Automatic Update and Completion of Occluded Regions for Accurate 3D

Urban Cartography by combining Multiple Views and Multiple Passages

Ahmad Kamal Aijazi, Paul Checchin and Laurent Trassoudaine

Clermont Universit

e, Universit

e Blaise Pascal, UMR 6602, CNRS, 63177 Aubiere, France

Keywords:

3D Cartography, Lidar Data, Occlusions, Multiple Views, Multiple Passages.

Abstract:

Handling occlusions is one of the more difﬁcult challenges faced today in urban landscape analysis and carto-

graphy. In this paper, we successfully address this problem by using a new method in which multiple views

and multiple sessions or passages are used to complete occluded regions in a 3D cartographic map. Two 3D

point clouds, from different viewing angles, obtained in each passage are ﬁrst classiﬁed into two main object

classes: Permanent and Temporary (which contains both Temporarily static and Mobile objects) using infer-

ence based on basic reasoning. All these Temporary objects, considered as occluding objects, are removed

from the scene leaving behind two perforated 3D point clouds of the cartography. These two perforated point

clouds from the same passage are then combined together to ﬁll in some of the holes and form a uniﬁed per-

forated 3D point cloud of the cartography. This uniﬁed perforated 3D point cloud is then updated by similar

subsequent perforated point clouds, obtained on different days and hours of the day, ﬁlling in the remaining

holes and completing the missing features/regions of the urban cartography. This automatic method ensures

that the resulting 3D point cloud of the cartography is most accurate containing only the exact and actual

permanent features/regions. Special update and reset functions are added to increase the robustness of the

method. The method is evaluated on a standard data set to demonstrate its efﬁcacy and prowess.

1 INTRODUCTION

In the recent past, 3D urban reconstruction and Vi-

sualisation of cities have become a hot topic of re-

search in the scientiﬁc community. Several geograph-

ical navigators like Google Streetmap Viewer, Mi-

crosoft Visual Earth or Geoportail provide real-like

representation of the terrain and ground based mod-

els, created by mobile terrestrial image acquisition

techniques. However, in urban environments the qua-

lity of data acquired by these hybrid terrestrial vehi-

cles is widely hampered by the presence of temporary

stationary and dynamic objects (pedestrians, cars, etc)

in the scene. As a result there is a problem of oc-

clusion of regions. Moving objects or certain tempo-

rary stationed objects (parked cars, trafﬁc, pedestrian

etc) present in the area hide certain zones of the ur-

ban landscape (buildings, road sides etc.). This pa-

per presents a new method for handling occlusions

and moreover automatic update and completion of oc-

cluded regions for accurate 3D urban cartography ex-

ploiting the concept of multiple views and multiple

passages.

In order to solve the problem of occlusions, diff-

erent techniques have been employed building on the

advances made in the work on texture synthesis. A

technique of inpainting based on the patch exemplar-

based technique was presented by (Criminisi et al.,

2004). This approach was extended by (Wang et al.,

2008) to also infer depth from stereo pairs. A sim-

ilar method was used by (Engels et al., 2011) for

3D data, but occlusions were found automatically us-

ing object-speciﬁc detectors making it more suitable

for larger data sets. In the context of building fa-

cades and urban reconstruction, increased contextual

knowledge is available by assuming structure’s pla-

narity and repetition of features such as ﬂoors, win-

dows, etc. Building models are reconstructed by de-

tecting ﬂoors and estimating building height in the

work of (Konushin and Vezhnevets, 2007). Occlu-

sions are removed by cloning upper ﬂoors and pro-

pagating them downward. A method relying on LI-

DAR point cloud to ﬁnd and remove occlusions by

combining image fusion and inpainting is presented

by (Benitez et al., 2010). (Xiao et al., 2009) seman-

tically segment street-side scenes into several classes,

including vegetation and vehicles, but do not actively

ﬁll in missing data. Instead, they rely on the missing

405

Kamal Aijazi A., Checchin P. and Trassoudaine L. (2013).

Automatic Update and Completion of Occluded Regions for Accurate 3D Urban Cartography by combining Multiple Views and Multiple Passages.

In Proceedings of the 2nd International Conference on Pattern Recognition Applications and Methods, pages 405-412

DOI: 10.5220/0004182704050412

 SciTePress

information being available from other views.

A method of aligning multiple scans from vari-

ous viewpoints to ensure the 3D scene model com-

pleteness for complex and unstructured underground

environments is discussed by (Craciun et al., 2010).

A technique for extracting features from urban buil-

dings by fusing camera and lidar data is presented by

(Becker and Haala, 2007) but it also fails to address

this problem. (Frueh et al., 2004) proposed a method

in which the point cloud is used to generate a 3D mesh

that is then classiﬁed as foreground or background.

Large holes in the background layer, caused by oc-

clusions from foreground layer objects, are then ﬁlled

by planner or horizontal interpolation. However, such

an approach may result in false features in case of in-

sufﬁcient repetitions or lack of symmetry (Li et al.,

2011). In our work, we aim to resolve this problem

by using a new approach combining both multiple

views and multiple passages such that ﬁrst multiple

views are exploited to ﬁll in some of the holes in the

3D point cloud followed by the use of multiple scans

of the same environment obtained at different days

and hours of the day to ﬁll in the remaining holes and

completing the missing regions of the urban cartogra-

phy. This ensures that the resulting 3D point cloud

of the cartography is most accurate containing only

the exact and actual permanent features/regions. An

overview of the method is presented in Algorithm 1.

2 3D SCAN REGISTRATION

Different features and robust landmarks extracted

from 3D images as points of interest and as refer-

ences for image mapping and scan registration have

commonly been used for different multi-sessional

SLAM (Simultaneous Localisation And Mapping) al-

gorithms. This approach works well in simple repeti-

tive paths. But some more complex situations can be

found in urban environments where the selected fea-

tures/regions can be occluded. When the data acqui-

ring vehicle enters from different directions, then the

path is not repetitive. As a result, the selected fea-

tures/regions may not be readily visible, etc. Thus,

in order to cater for this problem, the method of di-

rect geo-referencing of 3D LiDAR points is found

most suitable in our case. The method uses integrated

GPS/IMU data to directly orient laser data from its

local reference frame to the mapping reference frame

(WGS84).The advantage of using this method is that

the transformation between the local laser reference

frame and the mapping reference frame is known at

any given moment (as long as the laser is synchro-

nized), independently if the laser is collecting data

(a)

(b)

(c)

Figure 1: (a) The Vis Center’s LiDAR Truck. (b) Optech

LiDAR/GPS system along with IMU mounted on a rigid

frame. (c) The different viewing angles of the mounted Li-

DAR systems.

in a static mode or in kinematic mode. Thus, the

laser can be used as a pushbroom sensor sweeping the

scene with proﬁles while ﬁxing the scan angles as the

vehicle moves.

The data that we have used to evaluate our work

are the dynamic data set of the 3D Urban Data Chal-

lenge 2011, which contains dynamic scenes from

downtown Lexington, Kentucky, USA obtained from

the Vis Center’s (University of Kentucky) LiDAR

Truck containing two Optech LiDAR sensor heads

(high scan frequency up to 200 Hz), a GPS, an iner-

tial measurement unit and a spherical digital camera

as shown in Figure 1.

3 CLASSIFICATION OF 3D

URBAN ENVIRONMENT

We classify the urban environment into 2 main ca-

tegories: Permanent objects and Temporary objects.

In order to achieve this, the 3D point cloud is ﬁrst

segmented into objects which are then classiﬁed into

basic object classes. Once classiﬁed into these basic

classes, they are then grouped under one of the 2 men-

tioned categories. Although several methods have

been proposed for the classiﬁcation of urban environ-

ments, we have used one of the most recent meth-

ods (Aijazi et al., 2012) for this task. This method

presents a super-voxel based approach in which the

3D urban point cloud is ﬁrst segmented into vox-

els and then converted into super-voxels. These are

then clustered together using an efﬁcient Link-Chain

method to form objects. These objects are then clas-

siﬁed using local descriptors and geometrical features

into 6 main classes: {Road, Building, Car, Pole, Tree,

Unclassiﬁed}. The salient features of this method are

data reduction, efﬁciency and simplicity of approach.

The 3D point cloud obtained from each of the two

mounted LiDAR sensors is divided into the 6 object

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

406

Algorithm 1: Proposed method.

Input: 3D urban point clouds for passage number n

1: Classify 3D urban point cloud obtained from each of

the two sensors into 6 groups: {Road, Building, Car,

Pole, Tree, Unclassiﬁed}

2: Further classify the objects as: {Permanent, Temporary

}

3: Separate out Temporary objects leaving behind two

perforated 3D point clouds

4: Combine the two 3D point clouds of the two sensors

(different viewing angles) to ﬁll in some of the holes

and obtain a uniﬁed perforated 3D point cloud P(n

)

5: Store Temporary objects in a register R(n

)

6: Match and compare P(n

) with P(n

−1) to ﬁll in the

remaining holes and complete 3D cartography

Update:

7: Compare Temporary objects in R(n

) with R(n

−1)

8: Upgrade Temporary objects in R(n

) if they are re-

peated in n

update

number of passages as Permanent and

add in P(n

)

Reset:

9: Compare the skyline in P(n

) with that of P(n

−1) to

calculate 3D error difference

10: If same error difference re-occurs in n

reset

number of

passages then Reset the modiﬁed part of the building

in P(n

) with that in the recently acquired point cloud

(perforated)

11: Update and Store R(n

)

12: Store point cloud P(n

)

13: R(n

−1) ← R(n

) and P(n

−1) ← P(n

)

14: return P(n

)

classes using this method. These classiﬁed objects

are then further classiﬁed using inference based on

their basic characteristics (like roads, buildings, trees

and poles cannot move whereas cars and pedestrians

can be either Temporarily static or Mobile) into two

classes: {Permanent , Temporary}. The classiﬁcation

chart as per our inference is presented in Table 1.

Table 1: Object classiﬁcation chart.

Object type Permanent Temporary (Static or Mobile)

Road x

Building x

Tree x

Pole x

Car x

Pedestrian x

Unclassiﬁed x

Once the objects present in the urban scene are

classiﬁed into these two main classes, in each pas-

sage, the objects classiﬁed as Temporary are sepa-

rated from the scene (for each lidar) for each passage

to obtain perforated point clouds. This perforation

is due to occlusions caused by the temporarily static

and mobile objects in the scene. These two perfo-

rated point clouds (from different viewing angles) are

merged together to form a uniﬁed 3D point cloud ﬁl-

ling in some of the holes in the process.These uniﬁed

perforated 3D point clouds of the same place obtained

via a single passage on different days and at different

times are then combined together to complete the 3D

cartography as discussed in the following sections.

As the nature of the unclassiﬁed objects is not

known, they are considered temporary by default be-

cause all the objects classiﬁed as temporary are com-

pared in update phase. If the same objects belonging

to this class are found in repeated passages, they are

then upgraded as Permanent objects in the update

phase discussed in Section 5.2 and are considered part

of the 3D cartography.

4 COMBINING MULTIPLE

VIEWS

The two 3D point clouds obtained in the same pas-

sage are matched and merged together. This not only

helps in ﬁlling some of the holes but also completes

the 3D cartography (building fac¸ades, etc.) due to the

different viewing angles of the two sensors (from now

on referred to as S −01 and S −02) as shown in Fi-

gure 1(c). This conﬁguration of the LiDAR sensors

is very common for this type of sensors and is used

for acquiring detailed 3D data along both sides of the

road. We see that this conﬁguration of the lidar sen-

sors not only ﬁll in holes due to static objects present

in the scene (see Figure 5) but also those caused by

the mobile objects as due to this difference in viewing

angle these sensors see the same point in the 3D scene

with a slight time difference. This time difference al-

lows a moving object to move such that the occluded

portion of the cartography due to this object in S −01

becomes evident in S −02 while the occluded portion

in S −02 was evident in S −01. Thus, combining the

two 3D point clouds help ﬁll up such holes in an at-

tempt to eliminate the effect of the moving object in

the scene as shown in Figure 5.

The two 3D point clouds are matched and merged

together using the new matching method introduced

in Section 5.1 to form a uniﬁed perforated 3D point

cloud. This method of fusing multiple view data to

ﬁll in holes may be effective in simple cases but in

more complex scenarios offered in the urban envi-

ronment relying on multiple views alone does not en-

tirely solve the problem of occlusions and a number

of holes remain due to several blind spots as shown

in Figure 5 (neighborhood-2). In Figure 5(c) (clas-

sic case of occlusion caused by a temporary static

object, i.e. a parked car, present in the scene) it can

be seen that even after combining the data from mul-

tiple views, some regions of the cartography remain

AutomaticUpdateandCompletionofOccludedRegionsforAccurate3DUrbanCartographybycombiningMultipleViews

andMultiplePassages

407

occluded such as parts of shop’s wall, lamppost, tree

and road side.

Similarly for a temporary dynamic object, i.e. a

moving car present in the urban scene, combining

the data from multiple views does not fully complete

all the occluded regions as seen in Figure 5(f) (some

parts of shop’s wall and road side remain occluded).

So, this technique of combining multiple views is

helpful for completion of certain occluded regions but

does not solve the problem completely. Thus, in order

to complete the remaining exact and actual occluded

regions we use a new method exploiting the concept

of multiple passages described in the next section.

5 EXPLOITING MULTIPLE

PASSAGES

The uniﬁed perforated 3D point clouds obtained for

each passage of any particular place are combined to-

gether to ﬁll in the remaining holes and complete the

3D urban cartography. These 3D point clouds are ﬁrst

registered and then the 3D points are matched point

by point, completing the missing regions.

5.1 3D Urban Point Cloud Matching

Each subsequent 3D point cloud is registered with the

former point cloud by using the ICP method (Besl and

McKay, 1992). This method is most suitable for this

task as the 3D data are already geo-referenced and

hence lie in close proximity. It is observed that the

major part of the 3D urban point clouds is composed

of building points which are also found to be most

consistent. Thus instead of applying the ICP method

to complete 3D point clouds, only the building points

are taken into account. First the proﬁle/envelope of

the buildings is extracted and then the ICP method

is applied, matching these boundaries to obtain the

transformation matrix. The outlines/envelopes of the

buildings are extracted using a sweep scan method.

As the bottom part of the building outline close to

the ground is often occluded and hence inconsistent

due to the presence of different objects in the scene,

only the boundary of the top half of the building out-

line is subjected to ICP as shown in Figure 2.

Once the transformation matrix (rotation matrix

and translation matrix T

) is found, the whole 3D

point cloud P(n

) is transformed into P

) and then

registered with the former P(n

−1) using (1).

) = R

(P(n

)) + T

(1)

In order to avoid redundant points a union of 3D

points belonging to the two registered point clouds

Figure 2: In red and blue color we have the building outline

obtained from ﬁrst passage. In green and black we have the

building outline obtained from the second passage. Only

top half of the two outlines in blue and black respectively is

subjected to ICP.

is performed. Each 3D point of the ﬁrst cloud is

matched with that of the second cloud. Two corres-

ponding points match if (2) is satisﬁed:

−p

≤

√

tol

(2)

where p

and p

(3 ×1 vectors) are point positions

in the two point clouds along X, Y and Z axes. e

tol

(3 ×1 vector) is equal to the inverse of the maximum

number of 3D points per cubic meter that is desired in

the 3D cartographic map. The matched 3D points are

considered as one point. This ensures that only the

missing points are added completing the perforated

3D point cloud.

5.2 Update Phase

With every new passage, where the urban cartogra-

phy is completed, the objects classiﬁed in each pas-

sage as Temporary are analyzed. If the same objects

belonging to this class are found at the same place in

repeated passages, they are then upgraded as Perma-

nent objects and are considered part of the 3D car-

tography (see Algorithm 1). They are then added to

the 3D cartography. Otherwise non-repetitive objects

are deleted from the update register R(n

). This num-

ber of allowable repetitions n

update

can be ﬁxed, based

upon the frequency of repetition, time of repetition,

etc. This not only allows gradual update of the 3D car-

tography but also accommodates unclassiﬁed objects

in the scene; for example, in case of neighborhood-1

(see Figure 7), a few unclassiﬁed dustbins/trashcans

were added to the cartography, after repetition in the

successive passages, during update phase.

The Temporary objects in each passage are com-

pared with the objects present in the update register

R using the equations (3), (4) & (5). Simple match-

ing based on constituting points, color and intensity is

enough. Let P

Obj

and Q

Obj

be the two set of points

belonging to an object n in the update register R and

the new passage respectively. This Obj

is considered

to be repeated if only, and only if, the following three

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

408

conditions are satisﬁed:

Card(PQ

Obj

)

Card(min[P

Obj

, Q

Obj

])

×100 ≥w

(3)

Obj

nR,G,B

−Q

Obj

nR,G,B

≤ 3

√

(4)

Obj

−Q

Obj

≤ 3

√

(5)

where PQ

Obj

= P

Obj

nX,Y,Z

∩Q

Obj

nX,Y,Z

. PQ

Obj

is the set

containing the matched points obtained by point wise

intersection of the 3D points in the two sets if, and

only if, the difference in the distance along X, Y & Z

axes between two points in the P

Obj

nX,Y,Z

and Q

Obj

nX,Y,Z

is ≤ 2 ×P

Here P

is taken as the measurement accuracy

of the LiDAR sensor (value can be obtained from

data sheet) and Card is the cardinal number function.

Obj

nX,Y,Z

and Q

Obj

nX,Y,Z

are the sets of the 3D coordi-

nates of the points of the object while P

Obj

nR,G,B

and

Obj

nR,G,B

are the mean R,G & B values of the object

in P and Q point clouds respectively. P

Obj

and Q

Obj

are the mean laser reﬂectance intensity values in P

and Q point clouds respectively. w

is the matching

weight equal to the allowable percentage of the object

points whose position matches the two point clouds.

is the color weight equal to the maximum vari-

ance of R,G & B values for P

Obj

and Q

Obj

. w

the intensity weight equal to the maximum variance

of intensity values for P

Obj

and Q

Obj

5.3 Reset Phase

In case of a building or some part of a building is de-

molished in the 3D cartography due to renovation or

reconstruction then the 3D point cloud is reset with

the most recently acquired point cloud (perforated).

In such a case, it is assumed that the removal of com-

plete building or part of it causes a change in the sky-

line (or top part of the proﬁle). This change is de-

tected by comparing the skyline in the most recently

acquired point cloud with that of the former. The

3D error difference of the two skylines is analyzed as

shown in Figure 4. If the same error difference (∆x, ∆y

and ∆z in length, width and height respectively are

greater than pre-deﬁned thresholds) re-occurs in sub-

sequent n

reset

number of passages then only the modi-

ﬁed part of the building in the 3D point cloud is reset

(column width and thickness equal to the error size

along respective axis and height equal to the building

height) with that in the recently acquired point cloud

(perforated). The proposed reset method was veriﬁed

by synthetically removing/demolishing different parts

of the buildings in the 3D cartography as shown in Fi-

gure 3. In Figure 4, the skylines in the two 3D point

clouds and the 3D error difference are shown.

(a)

(b)

Figure 3: (a) shows the building in the ﬁrst 3D point cloud.

In (b) demolished building in the subsequent 3D point cloud

is presented.

Figure 4: The skylines in the two 3D point clouds are re-

spectively shown in (a) and (b). In (c) the 3D error of the

two skylines is presented.

6 RESULTS

In order to validate our method, the dynamic data set

of the 3D Urban Data Challenge 2011

was used.

This data set contains 4 sets of the same dynamic

scenes of downtown Lexington, Kentucky, USA ob-

tained on different days and at different times. The

data set consists in 3D points coupled with correspon-

ding laser reﬂectance intensity values. As the corres-

ponding RGB values are not readily available, (3) was

not used. This did not have much impact on the results

as laser reﬂectance values is found to be more consis-

tent than RGB values in an urban environment as it is

more illumination invariant (Aijazi et al., 2012). The

results of our method applied to two different neigh-

borhoods are discussed in this paper. In Figure 7,

the detailed results for neighborhood-1 are presented.

The value of w

= 75 (i.e. if more than 75% of the

object’s 3D points in the two point clouds match),

http://www.vis.uky.edu/3dimpvt/

AutomaticUpdateandCompletionofOccludedRegionsforAccurate3DUrbanCartographybycombiningMultipleViews

andMultiplePassages

409

(a) 3D points from S −01 (b) 3D points from S −02 (c) Combined 3D point clouds

(d) 3D points from S −01

(e) 3D points from S −01

(f) Combined 3D point clouds

Figure 5: In (a), (b) and (c) the ﬁlling of holes due to the presence of a temporarily static (parked) car by multiple view

combination is presented. In (d), (e) and (f) the ﬁlling of holes due to the presence of a temporary (dynamic) car by com-

bining multiple views, is presented. It can be seen in (c) and (f) that even after combining data from multiple views, some

holes/missing regions still remain due to blind spots.

Table 2: Classiﬁcation results for permanent objects (F

Score).

Neighborhood

Passage Combining

1 2 3 4 multiple passages

1 0.943 0.977 0.958 0.959 0.981

2 0.979 0.975 0.981 0.984 0.992

tol

= (0.000125 0.000125 0.000125)

(in m

) and

update

= 4 (maximum number of passages possible in

our case) was used to evaluate our results.

6.1 Evaluation and Discussion

In order to obtain a more comprehensive analysis,

ﬁrstly, the classiﬁcation results were evaluated for

all passages individually and then for the proposed

method of multiple passages using F-measure (see

(6)) as described by (Hripcsak and Rothschild, 2005).

= (1 + β

)

p.r

(β

. p + r)

(6)

where p and r are the precision and recall respectively

and β is the weight constant. The classiﬁcation results

of Permanent object type for neighborhood 1 and 2

are presented in Table 2. Constituting the permanent

cartography, they are of most interest to us whereas

all other object types are removed from the ﬁnal 3D

point cloud. The value of β = 1 was used to obtain

a balanced F

score. The classiﬁed objects were con-

sidered as a percentage of their constituting points in

the 3D scene. The Table 2 shows that the proposed

method also improves the classiﬁcation results.

Thus the proposed technique, independent of the

initial classiﬁcation method used, ensures that the per-

manent objects in the 3D urban scene are well charac-

(a) (b)

Figure 6: (a) shows the generated 3D model of the buildings

according to dimensions. In (b) one of the camera images

of the building is presented.

terized/extracted for further processing and cartogra-

phy.

In order to evaluate the accuracy of the completed

permanent features/regions, we selected a corner buil-

ding in neighborhood-1. As no ground truth was read-

ily available, we generated the ground truth by cre-

ating a simpliﬁed 3D model of the building using a

standard CAD software as shown in Figure 6. The

3D points from the initial acquisition were used for

this purpose while the missing features/regions were

completed by horizontal and vertical interpolation,

exploiting the symmetry of the building design, and

conﬁrmed/matched with the images of the building,

from different viewing angles, acquired by the digi-

tal spherical camera mounted on the vehicle (see Fi-

gure 6). A number of features/regions including the

occluded zones were selected for comparison. The di-

mensions of these selected features updated and com-

pleted in the 3D point cloud obtained at each passage,

using our method, was then compared with their cor-

responding dimensions in the ground truth. The ave-

rage absolute errors in X , Y and Z (height) axes of

the available dimensions obtained in each passage are

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

410

(a) 3D point cloud P(n

): after 1

passage

(b) 3D point cloud P(n

): after 3

passage

): after 4

passage

(d) 3D point cloud P(n

): after update phase

Figure 7: (a) shows the initial point cloud related to urban

cartography, full of holes due to occlusions. Completion

of occluded and missing regions in the urban cartography,

exploiting the concept of multiple passages are presented in

(b) & (c). (d) shows the updated point cloud after update

phase.

Figure 8: Average absolute errors in X, Y and Z (height)

directions.

presented in Figure 8. These error values include both

registration and sensor measurement errors. Low ave-

rage absolute error values obtained for passage-1 are

due to the fact that part of these 3D points was used

for ground truth generation.

Generally, these fairly constant error values are

acceptable for most applications, it is observed that

the average absolute error values increase slightly in

subsequent passages due to the fact that registration

errors add up in every passage.

7 CONCLUSIONS

In this work, we present a new method for 3D urban

cartography that successfully addresses the difﬁcult

problem concerning the automatic update and com-

pletion of occluded regions in the urban environment

by combining both multiple views and multiple pas-

sages. The proposed method ensures that the result-

ing 3D point cloud of the cartography is most accu-

rate that it contains only the exact and actual perma-

nent features/regions. The evaluated results demon-

strate the technical prowess of the proposed method

which can be easily integrated in different commercial

and non-commercial applications pertaining to urban

landscape modeling and cartography that need to up-

date their database frequently.

ACKNOWLEDGEMENTS

This work is supported by the French national re-

search agency (ANR CONTINT iSpace&Time –

ANR-10-CONT-23) and by “le Conseil G

eral de

l’Allier”. The authors would like to thank Prof.

Ruigang Yang for providing access to the 3D Urban

Data Challenge data set.

REFERENCES

Aijazi, A. K., Checchin, P., and Trassoudaine, L. (2012).

Classiﬁcation of 3D Urban Scenes – A Voxel Based

Approach. In Int. Conf. on Pattern Recognition, Ap-

plications and Methods, Vilamoura, Portugal.

Becker, S. and Haala, N. (2007). Combined Feature Extrac-

tion for Fac¸ade Reconstruction. In ISPRS Workshop

on Laser Scanning and SilviLaser, page 44, Espoo,

Finland.

Benitez, S., Denis, E., and Baillard, C. (2010). Automatic

Production of Occlusion-Free Rectiﬁed Fac¸ade Tex-

tures using Vehicle-Based Imagery. In Photogrammet-

ric Computer Vision and Image Analysis, page A:275.

Besl, P. and McKay, N. (1992). A Method for Registration

of 3-D Shapes. In IEEE Trans. on Pattern Analysis

and Machine Intelligence, volume 14(2), pages 239–

256, USA.

Craciun, D., Paparoditis, N., and Schmitt, F. (2010). Multi-

view scans alignment for 3D spherical mosaicing in

large-scale unstructured environments. Computer Vi-

sion and Image Understanding, 114(11):1248 – 1263.

AutomaticUpdateandCompletionofOccludedRegionsforAccurate3DUrbanCartographybycombiningMultipleViews

andMultiplePassages

411

Criminisi, A., P

erez, P., and Toyama, K. (2004). Region ﬁl-

ling and object removal by exemplar-based inpainting.

IEEE Trans. Image Processing, 13(9):1200–1212.

Engels, C., Tingdahl, D., Vercruysse, M., Tuytelaars, T.,

Sahli, H., and Van Gool, L. (2011). Automatic oc-

clusion removal from fac¸ades for 3D urban recons-

truction. Advanced Concepts for Intellig. Vision Sys-

tems, pages 681–692.

Frueh, C., Sammon, R., and Zakhor, A. (2004). Auto-

mated texture mapping of 3D city models with oblique

aerial imagery. In Proc. of 2nd Int. Symp. on 3D Data

Processing, Visualization and Transmission (3DPVT),

pages 396–403. IEEE.

Hripcsak, G. and Rothschild, A. S. (2005). Agreement, the

F-Measure, and Reliability in Information Retrieval.

Journal of American Medical Iinformatics Associa-

tion, 12(3):296–298.

Konushin, V. and Vezhnevets, V. (2007). Automatic buil-

ding texture completion. In Proc. Graphicon’2007,

Moscow, Russia.

Li, Y., Zheng, Q., Sharf, A., Cohen-Or, D., Chen, B., and

Mitra, N. J. (2011). 2D-3D Fusion for Layer Decom-

position of Urban Facades. In IEEE Int. Conf. on

Computer Vision (ICCV), Barcelona, Spain.

Wang, L., Jin, H., Yang, R., and Gong, M. (2008). Stereo-

scopic inpainting: Joint color and depth completion

from stereo images. In IEEE Conf. on Computer Vi-

sion and Pattern Recognition (CVPR), pages 1–8.

Xiao, J., Fang, T., Zhao, P., Lhuillier, M., and Quan, L.

(2009). Image-based street-side city modeling. ACM

Trans. Graph., 28:114:1–114:12.

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

412