Disocclusion of 3D LiDAR point clouds using range

images
Pierre Biasutti, Jean-François Aujol, Mathieu Brédif, Aurélie Bugeau

To cite this version:

Pierre Biasutti, Jean-François Aujol, Mathieu Brédif, Aurélie Bugeau. Disocclusion of 3D LiDAR point
clouds using range images. City Models, Roads and Traffic workshop (CMRT), Jun 2017, Hannover,
Germany. pp.75 - 82, �10.5194/isprs-annals-IV-1-W1-75-2017�. �hal-01522366�

HAL Id: hal-01522366

https://hal.archives-ouvertes.fr/hal-01522366
Submitted on 14 May 2017

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 DISOCCLUSION OF 3D LIDAR POINT CLOUDS USING RANGE
IMAGES
P. Biasuttia, b , J-F. Aujola , M. Brédifc , A. Bugeaub
a
Université de Bordeaux, IMB, CNRS UMR 5251, INP, 33400 Talence, France.
b
Université de Bordeaux, LaBRI, CNRS UMR 5800, 33400 Talence, France.
c
Université Paris-Est, LASTIG MATIS, IGN, ENSG, F-94160 Saint-Mandé, France.
{pierre.biasutti, aurelie.bugeau}@labri.fr, [email protected], [email protected]

Abstract in several ways, such as lidar point clouds colored by optical

images or lidar depth maps aligned with optical images.
This paper proposes a novel framework for the disocclusion Although these systems lead to very complete 3D mapping
of mobile objects in 3D LiDAR scenes aquired via street- of urban scenes by capturing optical and 3D details (pave-
based Mobile Mapping Systems (MMS). Most of the existing ments, walls, trees, etc.), they often acquire mobile objects
lines of research tackle this problem directly in the 3D space. that are not persistent to the scene. This often happens
This work promotes an alternative approach by using a 2D in urban environments with objects such as cars, pedestri-
range image representation of the 3D point cloud, taking ans, traffic cones, etc. As LiDAR sensors cannot penetrate
advantage of the fact that the problem of disocclusion has through opaque objects, those mobile objects cast shadows
been intensively studied in the 2D image processing commu- behind them where no point has been acquired (Figure 1,
nity over the past decade. First, the point cloud is turned left). Therefore, merging optical data with the point cloud
into a 2D range image by exploiting the sensor’s topology. can be ambiguous as the point cloud might represent ob-
Using the range image, a semi-automatic segmentation pro- jects that are not present in the optical image. Moreover,
cedure based on depth histograms is performed in order to these shadows are also largely visible when the point cloud
select the occluding object to be removed. A variational is not viewed from the original acquisition point of view.
image inpainting technique is then used to reconstruct the This might end up being distracting and confusing for visu-
area occluded by that object. Finally, the range image is alization. Thus, the segmentation of mobile objects and the
unprojected as a 3D point cloud. Experiments on real data reconstruction of their background remain a strategic issue
prove the effectiveness of this procedure both in terms of in order to improve the understability of urban 3D scans.
accuracy and speed. We refer to this problem as disocclusion in the rest of the
paper.
1 INTRODUCTION In real applicative contexts, we acknowledge that the dis-
occluded regions might cause a veracity issue for end-users
Over the past decade, street-based Mobile Mapping Sys- so that the reconstruction masks should be kept in the meta-
tems (MMS) have encountered a large success as the on- data. Using these masks, further processing steps may then
board 3D sensors are able to map full urban environments choose to process differently disoccluded regions (e.g. kept
with a very high accuracy. These systems are now widely for artefact-free visualizations or discarded during object ex-
used for various applications from urban surveying to city tractions).
modeling [27, 19, 14, 18, 17]. We argue that working on simplified representations of
Several systems have been proposed in order to perform the point cloud, especially range images, enables specific
these acquisitions. They mostly consist in optical cameras, problems such as disocclusion to be solved not only using
3D LiDAR sensor and GPS combined with Inertial Measure- traditional 3D techniques but also using techniques brought
ment Unit (IMU), built on a vehicle for mobility purposes by other communities (image processing in our case).
[22, 16]. They provide multi-modal data that can be merged In this work, we aim at presenting a novel framework for
the fast and efficient disocclusion of LiDAR point clouds.

Figure 1: One result of our proposed method. (left) original point cloud, (center) segmentation, (right) disocclusion.

1
Our first contribution is to provide a fast segmentation tech- leads to a fast segmentation but it might fail when the scale
nique for dense and sparse point clouds to extract full objects of the objects in the scene is too different. Another simpli-
from the scene by leveraging the implicit range image topol- fied model of the point cloud is presented by [31]. The au-
ogy (Figure 1, center). A second contribution is to introduce thors take advantage of the implicit topology of the sensor
a fast and efficient variational method for the disocclusion to represent the point cloud as a 2-dimensional range im-
of a point cloud using range image representation while tak- age in order to segment it before performing classification.
ing advantage of an horizontal prior without any knowledge The segmentation is done through a graph-based method
of the color or texture of the represented objects (Figure 1, as the notion of neighborhood is easily computable on a
right). 2D image. Although the provided segmentation algorithm
The paper is organized as follows: after a review on related is fast, it suffers from the same issues as geometry-based
state-of-the-art, we detail how the point cloud is turned into algorithms such as over-segmentation or incoherent segmen-
a range image. In section 3, both segmentation and disoc- tation. Moreover, all those categories of segmentation tech-
clusion aspects of the framework are explained. We then niques are not able to treat efficiently both dense and sparse
validate these approaches on different urban LiDAR data. LiDAR point clouds e.g. point clouds aquired with high or
Finally a conclusion and an opening are drawn. low sampling rates compared to the real-world feature sizes.
In this paper, we present a novel simplified model for seg-
mentation based on histograms of depth in range images.
2 Related Works
The growing interest for MMS over the past decade has lead 2.2 Disocclusion
to many works and contributions for solving problems of Disocclusion of a scene has only been scarcely investigated
segmentation and disocclusion. In this part, we present a for 3D point clouds [28, 24, 2]. These methods mostly work
state-of-the-art on both segmentation and disocclusion. on complete point clouds rather than LiDAR point clouds.
This task, also referred to as inpainting, has been much
2.1 Point cloud segmentation more studied in the image processing community. Over the
past decades, various approaches have emerged to solve the
The problem of point cloud segmentation has been exten- problem in different manners. Patch-based methods such as
sively addressed in the past years. Three types of methods [10] (and more recently [8, 21]) have proven their strengths.
have emerged: geometry-based techniques, statistical tech- They have been extended for RGB-D images [7] and to Li-
niques and techniques based on simplified representations of DAR point clouds [13] by considering an implicit topology
the point cloud. in the point cloud. Variational approaches represent an-
other type of inpainting algorithms [9, 5, 29, 3]. They have
Geometry-based segmentation The first well-known been extended to RGB-D images by taking advantage of the
method in this category is region-growing where the point bi-modality of the data [15, 4]. Even if the results of the dis-
cloud is segmented into various geometric shapes based on occlusion are quite satisfying, those models require the point
the neighboring area of each point [20]. Later, techniques cloud to have color information as well as the 3D data. In
that aim at fitting primitives (cones, spheres, planes, cubes this work, we introduce an improvement to a variational
...) in the point cloud using RANSAC [26] have been pro- disocclusion technique by taking advantage of a horizontal
posed. Others look for smooth surfaces [25]. Although those prior.
methods do not need any prior about the number of objects,
they often suffer from over-segmenting the scene and as a re-
sult objects are segmented in several parts. 3 Methodology
The main steps of the proposed framework, from the raw
Statistical segmentation The methods in this category point cloud to the final result, are described in Figure 2.
analyze the point cloud characteristics [12, 30, 6]. They con- We detail each of these steps in this section.
sider different properties of the PCA of the neighborhood of
each point in order to perform a semantic segmentation. It
leads to a good separation of points that belongs to static 3.1 Range maps and sensor topology
and mobile objects, but not to the distinction between dif- The key point of the proposed approach is to work on a
ferent objects of the same class. simplified representation of the point cloud known as a 2D
range map. The acquired dataset simply consisted in a map-
Simplified model for segmentation MMS LiDAR ping of the scene, the range map is obtained using the im-
point clouds typically represent massive amounts of unor- plicit topology of the sensor. The fact that most raw Li-
ganized data that are difficult to handle, different segmen- DAR acquisitions offer an intrinsic 2D sensor topology is
tation approaches based on a simplified representation of the rarely considered. Namely, LiDAR points may obviously
point cloud have been proposed. [23] proposes a method in be ordered along scanlines, yielding the first dimension of
which the point cloud is first turned into a set of voxels the sensor topology, linking each LiDAR pulse to the im-
which are then merged using a variant of the SLIC algo- mediately preceding and succeeding pulses within the same
rithm for super-pixels in 2D images [1]. This representation scanline. For most LiDAR devices, one can also order the

2
 a. b.

Figure 4: Result of the histogram segmentation using [11].

(a) segmented histogram (bins of 50cm), (b) result in the
range image using the same colors.
Figure 2: Overview of the proposed framework.
purposes, as it has the added advantage of providing pulse
neighborhoods that are reasonably local both in terms of
space and time, thus being robust to misregistrations, and
being very efficient to handle (constant time access to neigh-
bors). Moreover, as LiDAR sensor designs evolve to higher
sampling rates within and/or across scanlines, the sensor
topology will better approximate spatio-temporal neighbor-
hoods, even in the case of mobile acquisitions.
We argue that raw LiDAR datasets generally contain all
the information (scanline ordering, pulses with no echo,
number of points per turn...) to enable a constant-time
access to a well-defined implicit sensor topology. However
it sometimes occurs that the dataset received further pro-
cessings (points were reordered or filtered, or pulses with
no return were discarded). Therefore, the sensor topology
Figure 3: Example of a point cloud from the KITTI database may only be approximated using auxilliary point attributes
(top) turned into a range image (bottom). Note that the (time, θ, φ, fiber id...) and guesses about acquisition settings
black area in (b) corresponds to pulses with no returns. (e.g. guessing approximate ∆time values between successive
pulse emissions).
In the following sections, the range image is denoted uR .
consecutive scanlines so as to consider a second dimension of
the sensor topology across the scanlines. 2D LiDAR sensors
(i.e. : featuring a single simultaneous scanline acquisition)
3.2 Point cloud segmentation
generally send an almost constant number H of pulses per We now propose a segmentation technique based on range
scanline (or per turn for 360 degree 2D LiDARs), so that histograms. For the sake of simplicity, we assume that the
range measurements may be organized in an image of size ground is relatively flat and remove ground points by plane
W × H, where W is the number of consecutive scanlines and fitting.
thus a temporal dimension. 3D LiDAR sensors are based on Instead of segmenting the whole range image uR directly,
multiple simultaneous scanline acquisitions (e.g. H = 64) we first split this image in S sub-windows uR s ,s = 1...S
such that scanlines may be stacked horizontally to form an of size Ws × H along the horizontal axis. For each uR s, a
image, as illustrated in Figure 3. depth histogram hs of B bins is built. This histogram is au-
Whereas LiDAR pulses are emitted somewhat regularly, tomatically segmented into Cs classes using the a-contrario
many pulses yield no range measurements due, for instance, technique presented in [11]. This technique presents the ad-
to reflective surfaces, absorption or absence of target objects vantage of segmenting a 1D-histogram without any prior
(e.g. in the sky direction). Therefore the sensor topology is assumption, e.g. the underlying density function or the
only a relevant approximation for emitted pulses but not for number of objects. Moreover, it aims at segmenting the
echo returns, such that the range image is sparse with unde- histogram following an accurate definition of an admissible
fined values where pulses measured no echoes. Considering segmentation, preventing over and under segmentation from
multi-echo datasets as a multilayer depth image is beyond appearing. Examples of segmented histograms are given in
the scope of this paper. Figure 4.
The sensor topology only provides an approximation of Once the histogram of successive sub-images have been
the immediate 3D point neighborhoods, especially if the sen- segmented, we merge together the corresponding classes by
sor moves or turns rapidly compared to its sensing rate. We checking the distance between each of their centroids. Let
argue however that this approximation is sufficient for most us define the centroid of the ith class Csi in the histogram hs

3
 define the direction in which the diffusion should be done to
respect this prior. Note that most of MLS systems provide
georeferenced coordinates of each point that can be used to
define ~η .
We aim at extending the level lines of u along ~η . This can
be expressed as h∇u, ~η i = 0. Therefore, we define the energy
F (u) = 21 (h∇u, ~η i)2 . The disocclusion is then computed
as a solution of the minimization problem inf u F (u). The
2
Figure 5: Example of point cloud segmentation using our gradient of this energy is given by ∇F (u) = −h(∇ u)~η , ~η i =
model on various scenes. −uη~η~ , where uη~η~ stands for the second order derivative of
u with respect to ~η and ∇2 u for the Hessian matrix. The
minimization of F can be done by gradient descent. If we
of the sub-image uR
s as follows: cast it into a continuous framework, we end up with the
P following equation to solve our disocclusion problem:
b × hs (b)  ∂u
b∈Csi
Csi = P (1) ∂t − uη
~η~ = 0 in Ω × (0, T )
(4)
hs (b) u(0, x) = u0 (x) in Ω
b∈Csi
using previously mentioned notations. We recall that ∆u =
where b are all bins belonging to class Csi .
The distance uη~η~ +uη~T η~T , where ~η T stands for a unitary vector orthogonal
between two classes Csi and Crj , of two consecutive windows to ~η . Thus, Equation (4) can be seen as an adaptation the
can be defined as follows: Gaussian diffusion equation (3) to respect the diffusion prior
in the direction ~η . Figure 7 shows a comparison between the
d(Csi , Crj ) = |Csi − Crj | (2) original Gaussian diffusion algorithm and our modification.
The Gaussian diffusion leads to an over-smoothing of the
Finally, we can set a threshold such that if d(Csi , Crj ) ≤ scene, creating an aberrant surface whereas our modification
τ , classes Csi and Crj should be merged. Results of this provides a result that is more plausible.
segmentation procedure can be found in Figure 5. We argue The equation proposed in (4) can be solved iteratively.
that the choice of Ws , B and τ mostly depends on the type The number of iterations simply depends on the size of the
of data that is being treated (sparse or dense). For sparse area that needs to be filled in.
point clouds, B has to remain small (e.g. 50) whereas for
dense point clouds, this value can be increased (e.g. 200).
3.4 Range image to 3D point cloud
In practice, we found out that good segmentations may be
obtained on various kind of data by setting Ws = 0.5×B and After the segmentation and the disocclusion, we need to turn
τ = 0.2 × B. Note that the windows are not required to be the range image back to the 3D space. For every point pi
overlapping in most cases, but for very sparse point clouds, of the original point cloud, we define poi and pei respectively
an overlap of 10% is enough to reach good segmentation. the point of emission and the point of echo of pi . We denote
dorig (pi ) the original range of pi and drec (pi ) its range after
3.3 Disocclusion disocclusion. The new coordinates pfi inal of each point can
be obtained using the following formula:
The segmentation technique introduced above provides
drec (pi )
masks for the objects that require disocclusion. As men- pfi inal = poi + (pei − poi ) × (5)
tioned in the beginning, we propose a variational approach d orig (pi )
to the problem of disocclusion of the point cloud. The Gaus- The range image can then be easily turned back to a 3D
sian diffusion algorithm provides a very simple algorithm for point cloud while including the disocclusion.
the disocclusion of objects in 2D images by solving partial
differential equations. This technique is defined as follows:
 ∂u
∂t − ∆u = 0 in Ω × (0, T ) (3)
u(0, x) = u0 (x) in Ω

having u an image defined on Ω, t being a time range and

∆ the Laplacian operator. As the diffusion is performed
in every direction, the result of this algorithm is often very
smooth. Therefore, the result in 3D lacks of coherence as
shown in Figure 7.b.
In this work, we assess that the structures that require
disocclusion are likely to evolve smoothly along the xW and
yW axis of the real world as defined in Figure 6. Therefore,
we set ~η for each pixel to be a unitary vector orthogonal to Figure 6: Definition of the different frames between the Li-
the projection of zW in the uR range image. This vector will DAR sensor (xL , yL , zL ) and the real world (xW , yW , zW ).

4
 4 Results
In this part, the results of the segmentation of various ob-
jects and the disocclusion of their background are detailed.

4.1 Sparse point cloud

A first result is shown in Figure 8. This result is obtained
for a sparse point cloud (≈ 106 pts) of the KITTI database
[16]. A pedestrian is segmented out of the scene using our
proposed segmentation technique. The segmentation result
is used as a mask for the disocclusion of its background using
our modified variational technique for disocclusion. Figure
8.a shows the original range image. In Figure 8.b, the dark
region corresponds to the result of the segmentation step on
the pedestrian. For practical purpose, a very small dilata-
tion is applied to the mask (radius of 2px in sensor topology)
to ensure that no outlier points (near the occluder’s silhou-
a. ette with low accuracy or on the occluder itself) bias the
reconstruction. Finally, Figure 8.c shows the range image
after the reconstruction. We can see that the disocclusion
performs very well as the pedestrian has completely disap-
peared and the result is visually plausible in the range image.
In this scene, ~η has a direction that is very close to the x
axis of the range image and the 3D point cloud is acquired
using a panoramic sensor. Therefore, the coherence of the
reconstruction can be checked by looking how the acquisition
lines are connected. Figure 9 shows the reconstruction of the
same scene in 3 dimensions. We can see that the acquisition
lines are properly retrieved after removing the pedestrian.
This result was generated in 4.9 seconds using Matlab on a
2.7GHz processor. Note that a similar analysis can be done
on the results presented in Figure 1.
b.

4.2 Dense point cloud

In this work, we aim at presenting a model that performs
well on both sparse and dense data. Figure 10 shows a result
of the disocclusion of a car in a dense point cloud. This
point cloud was acquired using the Stereopolis-II system [22]
and contains over 4.9 million points. In Figure 10.a, the
original point cloud is displayed with the color based on
the reflectance of the points for a better understanding of
the scene. Figure 10.b highlights the segmentation of the car

c.

Figure 7: Comparison between disocclusion algorithms. (a)

is the original point cloud (white points belong to the object
to be disoccluded), (b) the result after Gaussian diffusion
and (c) the result with our proposed algorithm. a. b. c.

Figure 8: Result of disocclusion on a pedestrian on the

KITTI database [16]. (a) is the original range image, (b)
the segmented pedestrian (dark), (c) the final disocclusion.
Depth scale is given in meters.

5
 a.
a.

b.

b.

Figure 9: Result of the disocclusion on a pedestrian in 3D.

(a) is the original mask highlighted in 3D, (b) is the final
reconstruction.

using our model, dilated to prevent aberrant points. Finally,

Figure 10.c depicts the result of the disocclusion of the car
using our method.
We can note that the car is perfectly removed from the c.
scene. It is replaced by the ground that could not have
been measured during the acquisition. Although the recon- Figure 10: Result of the disocclusion on a car in a dense
struction is satisfying, some gaps are left in the point cloud. point cloud. (a) is the original point cloud colorized with the
Indeed, in the data used for this example, pulse returns with reflectance, (b) is the segmentation of the car highlighted in
large deviation values were discarded. Therefore, the win- orange, (c) is the result of the disocclusion.
dows and the roof of the car are not present in the point
cloud before and after the reconstruction as no data is avail-
able. we computed the MAE (Mean Absolute Error) between the
ground truth and the reconstruction (where the occlusion
was simulated) using both Gaussian disocclusion and our
4.3 Quantitative analysis model. We recall that the MAE is expressed as follows:
To conclude this section, we perform a quantitative analysis 1 X
MAE(u1 , u2 ) = |u1 (i, j) − u2 (i, j)| (6)
of our disocclusion model on the KITTI dataset. The ex- N
i,j∈Ω
periment consists in removing areas of various point clouds
in order to reconstruct them using our model. Therefore, where u1 , u2 are images defined on Ω with N pixels. Ta-
the original point clouds can serve as ground truths. Note ble 1 sums up the result of our experiment. We can note
that areas are removed while taking care that no objects are
present in those locations. Indeed, this test aims at show-
ing how the disocclusion step behaves when reconstructing Table 1: Comparison of the average MAE (Mean Absolute
backgrounds of objects. The size of the removed areas cor- Error) on the reconstruction of occluded areas.
responds to an approximation of a pedestrian’s size at 8 Gaussian Proposed model
meters from the sensor in the range image (20 × 20px).
Average MAE (meters) 0.591 0.0279
The test was done on 20 point clouds in which an area
STD of MAEs 0.143 0.0232
was manually removed and then reconstructed. After that,

6
 sical imaging technique that takes the nature of the point
cloud into account (horizontality prior on the 3D embed-
ding), leading to better results. The segmentation step can
be done in streaming any time a new window is acquired,
leading to great speed improvement, constant memory pro-
cessing and the possibility of online processing during the
acquisition. Moreover, our model is designed to work semi-
automatically using very few parameters in reasonable com-
putational time. Finally, we have shown that this work per-
a. b. forms well in various cases, both on dense and sparse point
clouds.
Considering the range image derived from the sensor
topology enabled a simplified formulation of the problem
from having to determine an unknown number of 3D points
to estimating only the 1D depth in the ray directions of a
fixed set of range image pixels. Beyond simplifying drasti-
cally the search space, it also provides directly a reasonable
sampling pattern for the reconstructed point set.
Although the average results of the method are more than
c. d. acceptable, it can underperform in some specific cases. In-
deed, the segmentation step first relies on the good extrac-
tion of non-ground points, which can be tedious when the
data quality is low. Moreover, when the object that needs
to be removed from the scene is hiding complex shapes, the
disocclusion step can fail recovering all the details of the
background and the result ends up being too smooth. This
is likely to happen when disoccluding very large objects.
In the future, we will focus on improving the current
model to perform better reconstruction by taking into ac-
e. f. count the neighborhood of the background of the object to
remove either by using a variational method or by extending
Figure 11: Example of results obtained for the quantitative patch-based method.
experiment. (a) is the original point cloud (ground truth),
(b) the artificial occlusion in dark, (c) the disocclusion result
with the Gaussian diffusion, (d) the disocclusion using our 6 Acknowledgement
method, (e) the Absolute Difference of the ground truth
against the Gaussian diffusion, (f) the Absolute Difference J-F. Aujol is a member of Institut Universitaire de France.
of the ground truth against our method. Scales are given in This work was funded by the ANR GOTMI (ANR-16-CE33-
meters. 0010-01) grant.

that our method provides a great improvement compared

to the Gaussian disocclusion, with an average MAE lower References
than 3cm. This result is largely satisfying as most of the [1] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and
structures to reconstruct were situated from 12 to 25 meters S. Süsstrunk. SLIC superpixels compared to state-of-the-
away from the sensor. art superpixel methods. IEEE Trans. on Pattern Analysis
Figure 11 shows an example of disocclusion following this and Machine Intelligence, 34, 2012.
protocole. The result of our proposed model is visually [2] J. Becker, C. Stewart, and R. J. Radke. LIDAR inpaint-
very plausible whereas the Gaussian diffusion ends up over- ing from a single image. In IEEE Int. Conf. on Computer
smoothing the reconstructed range image which increases Vision, 2009.
the MAE.
[3] M. Bertalmio, G. Sapiro, V. Caselles, and C. Ballester. Im-
age inpainting. In ACM Comp. graphics and interactive
techniques, 2000.
5 Conclusion
[4] M. Bevilacqua, J-F. Aujol, P. Biasutti, M. Brédif, and
In this paper, we have proposed a novel approach for the seg- A. Bugeau. Joint inpainting of depth and reflectance with
mentation and the disocclusion of objects in 3D point clouds visibility estimation. ISPRS Journal of Photogrammetry and
acquired using MMS. This model takes advantage of a sim- Remote Sensing, 2017.
plified representation of the point cloud known as a range [5] K. Bredies, K. Kunisch, and T. Pock. Total generalized
image. We have also proposed an improvement of a clas- variation. SIAM Journal on Imaging Sciences, 3, 2010.

7
 [6] M. Brédif, B. Vallet, and B. Ferrand. Distributed [22] N. Paparoditis, J-P. Papelard, B. Cannelle, A. Devaux,
dimensionality-based rendering of LIDAR point clouds. In- B. Soheilian, N. David, and E. Houzay. Stereopolis II: A
ternational Archives of the Photogrammetry, Remote Sens- multi-purpose and multi-sensor 3D mobile mapping system
ing and Spatial Information Sciences, 40, 2015. for street visualisation and 3D metrology. Revue française
de photogrammétrie et de télédétection, 200, 2012.
[7] P. Buyssens, M. Daisy, D. Tschumperlé, and O. Lézoray.
Depth-aware patch-based image disocclusion for virtual view [23] J. Papon, A. Abramov, M. Schoeler, and F. Worgotter.
synthesis. In ACM SIGGRAPH, 2015. Voxel cloud connectivity segmentation-supervoxels for point
clouds. In IEEE Conf. on Computer Vision and Pattern
[8] P. Buyssens, M. Daisy, D. Tschumperlé, and O. Lézo-
Recognition, 2013.
ray. Exemplar-based inpainting: Technical review and new
heuristics for better geometric reconstructions. IEEE Trans. [24] S. Park, X. Guo, H. Shin, and H. Qin. Shape and appearance
on Image Processing, 24, 2015. repair for incomplete point surfaces. In IEEE Int. Conf. on
Computer Vision, volume 2, 2005.
[9] A. Chambolle and T. Pock. A first-order primal-dual al-
gorithm for convex problems with applications to imaging. [25] T. Rabbani, F. Van Den Heuvel, and G. Vosselmann. Seg-
Jour. of Math. Imag. and Vis., 40(1), 2011. mentation of point clouds using smoothness constraint. In-
ternational Archives of the Photogrammetry, Remote Sens-
[10] A. Criminisi, P. Pérez, and K. Toyama. Region filling and ing and Spatial Information Sciences, 36, 2006.
object removal by exemplar-based image inpainting. IEEE
Trans. on Image Processing, 13, 2004. [26] R. Schnabel, R. Wahl, and R. Klein. RANSAC based out-
of-core point-cloud shape detection for city-modeling. Pro-
[11] J. Delon, A. Desolneux, J-L. Lisani, and A. B. Petro. A ceedings of “Terrestrisches Laserscanning", 2007.
nonparametric approach for histogram segmentation. IEEE
Trans. on Image Processing, 16, 2007. [27] A. Serna and B. Marcotegui. Urban accessibility diagnosis
from mobile laser scanning data. ISPRS Journal of Pho-
[12] J. Demantke, C. Mallet, N. David, and B. Vallet. Dimension- togrammetry and Remote Sensing, 84, 2013.
ality based scale selection in 3D LIDAR point clouds. Inter-
[28] A. Sharf, M. Alexa, and D. Cohen-Or. Context-based sur-
national Archives of the Photogrammetry, Remote Sensing
face completion. ACM Trans. on Graphics, 23, 2004.
and Spatial Information Sciences, 38, 2011.
[29] J Weickert. Anisotropic diffusion in image processing, vol-
[13] D. Doria and R. Radke. Filling large holes in lidar data
ume 1. Teubner Stuttgart, 1998.
by inpainting depth gradients. In Int. Conf. on Pattern
Recognition, 2012. [30] M. Weinmann, B. Jutzi, S. Hinz, and C. Mallet. Semantic
point cloud interpretation based on optimal neighborhoods,
[14] S. El-Halawany, A. Moussa, D. Lichti, and N. El-Sheimy. relevant features and efficient classifiers. ISPRS Journal of
Detection of road curb from mobile terrestrial laser scan- Photogrammetry and Remote Sensing, 105, 2015.
ner point cloud. In Proceedings of the ISPRS Workshop on
Laserscanning, Calgary, Canada., volume 2931, 2011. [31] X. Zhu, H. Zhao, Y. Liu, Y. Zhao, and H. Zha. Segmentation
and classification of range image from an intelligent vehicle
[15] D. Ferstl, C. Reinbacher, R. Ranftl, M. Rüther, and in urban environment. In IEEE Trans. on Intelligent Robots
H. Bischof. Image guided depth upsampling using and Systems, 2010.
anisotropic total generalized variation. In IEEE Int. Conf.
on Computer Vision, 2013.
[16] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision
meets robotics: The KITTI dataset. International Journal
of Robotics Research, 2013.
[17] F. Goulette, F. Nashashibi, I. Abuhadrous, S. Ammoun, and
C. Laurgeau. An integrated on-board laser range sensing
system for on-the-way city and road modelling. Interna-
tional Archives of the Photogrammetry, Remote Sensing and
Spatial Information Sciences, 34, 2006.
[18] A. Hervieu and B. Soheilian. Semi-automatic
road/pavement modeling using mobile laser scanning.
ISPRS Annals of the Photogrammetry, Remote Sensing and
Spatial Information Sciences, 2, 2013.
[19] A. Hervieu, B. Soheilian, and M. Brédif. Road marking ex-
traction using a MODEL&DATA-DRIVEN Rj-Mcmc. IS-
PRS Annals of the Photogrammetry, Remote Sensing and
Spatial Information Sciences, 2, 2015.
[20] J. Huang and C-H. Menq. Automatic data segmentation for
geometric feature extraction from unorganized 3D coordi-
nate points. IEEE Trans. on Robotics and Automation, 17,
2001.
[21] L. Lorenzi, F. Melgani, and G. Mercier. Inpainting strategies
for reconstruction of missing data in VHR images. IEEE
Trans. on Geoscience and Remote Sensing Letters, 8, 2011.