Aerial 3D Building Detection and Modeling From Airborne Lidar Point Clouds
Aerial 3D Building Detection and Modeling From Airborne Lidar Point Clouds
Aerial 3D Building Detection and Modeling From Airborne Lidar Point Clouds
3, JUNE 2013
Abstract—A fast, completely automated method to create 3D available data source. Research on how to extract 3D infor-
watertight building models from airborne LiDAR point clouds is mation from ground or aerial imagery has been conducted for
presented. The proposed method analyzes the scene content and decades. Recent advances have enabled techniques to directly
produces multi-layer rooftops with complex boundaries and ver-
tical walls that connect rooftops to the ground. A graph cuts based capture 3D information over large scale areas. With the emer-
method is used to segment vegetative areas from the rest of scene gence of LiDAR (light detection and ranging) technology, a
content. The ground terrain and building rooftop patches are powerful 3D representation in the form of a point cloud can
then extracted utilizing our technique, the hierarchical Euclidean been created to assist in the generation of 3D scenes in a more
clustering. Our method adopts a “divide-and-conquer” strategy.
Once potential points on rooftops are segmented from terrain and
efficient and cost effective manner. Many modern techniques
vegetative areas, the whole scene is divided into individual pendent are developed relying on the input from LiDAR.
processing units which represent potential building footprints. For When collecting nadir-looking aerial data in an urban set-
each individual building region, significant features on the rooftop ting, the two scene components that dominate this data are the
are further detected using a specifically designed region growing tops of man-made structures (rooftops) and the background ter-
algorithm with smoothness constraint. Boundaries for all of these
features are refined in order to produce strict description. After rain. The detection and segmentation of these rooftops from
this refinement, mesh models could be generated using an existing the terrain is hence a crucial step in building extraction. Often
robust dual contouring method. times, the outlines of these building rooftops are not simple
Index Terms—3D, building, LiDAR, vegetation, graph cuts, re- geometric primitives, on the contrary, they can be quite com-
gion growing, modeling. plex making the model extraction process a challenging task. In
this research, the key contribution is the fully automatic work
flow that is able to exploit useful information to urban modeling
I. INTRODUCTION from airborne LiDAR data only, effectively and robustly con-
T HREE dimensional building reconstruction has been a ducting the task of scene classification without spectral knowl-
highly active research topic for years. There has been edge from optical image data and three dimensional building
an increasing demand in various applications such as urban extraction and its geometric modeling. The process removes
planning, virtual tourism, computer gaming, real-time emer- vegetations through a graph cuts based approach and detects
gency response, and robot navigation. Commercial applications the terrain and building footprints by an Euclidean clustering
like Google Earth and Apple Maps have already deployed technique and then produces simplified, 3D meshing building
3D building reconstruction techniques as a vital visualiza- models using solely LiDAR point cloud data for a large urban
tion component which has gained huge public acceptance. scene. One major challenging issue at this stage is how to ef-
Models are usually created by texture mapping both aerial and ficiently and accurately segment building regions from the rest
ground-based images onto 3D geometric models. Traditionally, of the background regions in the scene, particularly vegetation,
geometrical models are built up manually. There are some without the assistance of multi-spectral optical imagery. The
very capable, free tools like Google Sketch-Up, that allow the graph cuts based approach designed in this research does a fine
layperson to accomplish this, however, it does require a lot of job on the scene classification. In addition, an effective method
human effort to create just a single building. It remains a very for the extraction of rooftop features is presented, by utilizing
difficult and arduous task, especially when a large cityscape this method, a great amount of details on the rooftop are well
needs to be created. In the remote sensing community, there are maintained.
several data sources that are suitable as input to the building
II. RELATED WORK
reconstruction process. Imagery is the traditional and most
There has been a plethora of work conducted for urban mod-
eling from airborne LiDAR data, aerial images, or the combi-
Manuscript received September 28, 2012; revised February 12, 2013; ac-
cepted March 02, 2013. Date of publication May 14, 2013; date of current ver- nation. Hu et al. [1] described a few projects on the topic of
sion June 17, 2013. This work was supported by the U.S. Department of Energy large scale urban modeling. Musialski et al. [2] recently pro-
under BAA PDP08 Grant Number DE-AR52-07NA28115.
vided a more comprehensive overview of urban reconstruction
S. Sun and C. Salvaggio are with the Digital Imaging and Remote Sensing
Laboratory, Chester F. Carlson Center for Imaging Science, Rochester Insti- from different perspectives which are not restricted to aerial data
tute of Technology, Rochester, NY 14623 USA (e-mail: sxs4643@rit.edu; sal- inputs. The complexity and difficulty of this problem has been
vaggio@cis.rit.edu).
approached in many ways, but the synergistic use of muli-modal
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. datasets has become a prominent pedagogy in this research area.
Digital Object Identifier 10.1109/JSTARS.2013.2251457 Sirmacek et al. [3] introduced a novel and robust approach for
Fig. 1. Building detection and modeling work flow: classification identifies the terrain, trees and building footprints; significant planar features are detected on
each individual rooftop; boundaries of all parts are produced; finally, the models are reconstructed by a volumetric method.
city modeling and also provided a detailed assessment of the al- [16]. This approach produces building models with arbitrary
gorithm performance on different types of remote sensed data, rooftop shapes.
including both airborne and space-borne data. Lafarge et al. [17], [18] presented a novel and robust method
Using 2D image data only, Moons et al. [4] proposed a for modeling cities from point clouds. The algorithm is able
method to reconstruct 3D polyhedral models. This method to construct simultaneously buildings, trees and terrains. They
requires that accurate camera model information is readily used LiDAR point clouds for experiments but claiming it is not
available. Kim et al. [5] presented an approach to automati- restricted to LiDAR data inputs.
cally describe a complex rooftop from multiple images. Image In consideration of the combination of LiDAR and geo-
derived elevation data is used to assist feature matching. Once metrically uncorrected image data for urban modeling, some
all of the 3D line features are obtained, the rooftop hypoth- work has been done on automatic registration of aerial images
esis generation process introduces the next important issue, with LiDAR data. Ding et al. [19] and Lu et al. [20] both
time complexity. Nevatia et al. [6] introduced an automatic proposed registration method based on 3D feature detection
and interactive modeling method from aerial images, and and matching. Mastin et al. [21] introduced a novel idea for
Rau et al. [7] presented a Split-Merge-Shape algorithm for utilizing mutual information between the LiDAR and the 2D
3D building modeling in which an accurate scheme for 3D imagery, which analyzes the statistical dependency in scenes.
roof-edge measurements is proposed, however, these two ideas This information merger has advantages, however, finding
are semi-automatic, still requiring manual intervention. correspondences between these two different types of data
Haala et al. [8] point out “the difficulties of aerial image inter- automatically is often problematic.
pretation also motivated the increasing use of three-dimensional
point clouds from laser altimetry as an alternative data source”. III. METHODOLOGY OVERVIEW
Using LiDAR data only, Wang et al. [9] developed a In this paper, a fully automatic method is proposed (see Fig. 1)
Bayesian method to detect building footprints automatically that processes airborne LiDAR point cloud data for the purpose
from the LiDAR data. The point cloud has to be first seg- of building detection and modeling. The LiDAR data is a group
mented in to buildings, trees and grass as a pre-processing step. of unorganized discrete points in which each individual point
Verma et al. [10] presented a method to detect and construct has an x, y, and z value, plus the intensity value that represents
complex buildings using LiDAR. It makes no assumption the reflective properties of surface encountered. Our algorithm
about prior knowledge about the types of buildings, but it requires only LiDAR data as input with no prerequisite to ras-
doesn’t address how to handle the situation if the rooftop has terize this data. The output of our algorithm is a group of water-
multiple layers/planar surfaces. Dorninger et al. [11] proposed tight mesh models that could be used for various applications.
a comprehensive approach for automated determination of We follow a standardized pipeline and introduce novel method-
3D city models from LiDAR point cloud data, but their ap- ologies to improve accuracy and efficiency. The steps in the
proach involved an interactive initialization called the coarse workflow pipeline are:
selection of building regions. Poullis et al. [12] addressed a 1) Preparation: The input point cloud is first passed through
rapid modeling approach for large-scale area reconstruction by a noise filter (statistical outlier removal [22]) in order to ob-
using statistical considerations for segmenting the buildings. tain a less noisy data set. The point normals and curvatures
Zhou et al. [13] proposed a workflow for automatic building of all input points are estimated based on basic eigenanal-
reconstruction from airborne LiDAR data. The highlights of ysis.
this method are SVM-based vegetation detection, boundary 2) Scene Classification: A graph cuts based optimization
extraction, and automatic determination of principal directions. algorithm is introduced which examines the local distribu-
In their follow-on work, they mainly focused on the modeling tions of point normals. These distributions exhibit vastly
step in the workflow and extended the classic dual contouring different behavior with a strong dependence on which
of hermite data [14] to a 2.5D dual-contouring technique [15]. landcover category (vegetation of non-vegetation) a set
They later improved the approach by adding topology control of points belongs to. A hierarchical Euclidean clustering
1442 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 3, JUNE 2013
Fig. 2. 3D color plot according to the value of : reddish color indicates smaller values; greenish color indicates larger values.
method is then proposed to initially extract the terrain A. Examine Local Distribution of Normals
from the tree-excluded scene, followed by the extraction In order to explore the properties of the local distribution of
of rooftop patches. After this step, each building footprint normals, the normal of each point needs to be estimated. The es-
is successfully detected (see the second block of the flow- timation is calculated within a local neighborhood of the query
chart in Fig. 1). point. This local neighborhood can be defined by the spatial re-
3) Rooftop Features Detection: In order to describe each in- lationships of the points in terms of 3D point cloud processing.
dividual building rooftop with the best spatial detail pos- The points in the neighborhood should be sufficient to represent
sible, a region growing, segmentation method is developed a small surface patch for feature analysis. Let be a sample
with a smoothness constraint and curvature consistency to point in the original point cloud,
detect every significant feature on the rooftop (see the third be the set of points within a radius range of point , and is
block of the flowchart in Fig. 1). the centroid of all points in is the search radius which de-
4) Boundary Production: Points on the boundaries of all de- termines the size of the neighborhood. A common solution for
tected parts on the rooftop are generated by applying recti- finding the point normal is to solve for the primary eigenvectors
linear fitting see the fourth block of the flowchart in Fig. 1. of the covariance matrix of points in this neighborhood. The co-
5) Building modeling: A robust 2.5D dual contouring variance matrix is defined as
method is utilized to generate facetized, watertight
building models see the last block of the flowchart in (1)
Fig. 1.
Fig. 3. The example scene classification result (blue: buildings; green: trees; black: terrain). The trees are detected by the graph cuts optimization algorithm
introduced in Section IV.B. The terrain and the building footprints are separated by the Euclidean clustering method introduced in Section IV.C.
B. Graph Cuts Optimization In this approach, each point in the point cloud represents a
node of the graph. Each point is connected by its four nearest
In computer vision, a larger number of tasks address the neighbors. These connections are edges in the graph and can
problem of assigning a label to each pixel in a 2D image. For be weighted proportionally by distance. In the context of this
3D point cloud data, we would like to conduct the same task problem, the data term presents a penalty if a point with a small
and assign each three-dimensional point a label. A graph cut value (the variation of local distribution of normals) is en-
algorithm with -expansion move [24], [25] will be used for countered and is labeled as a tree, and vice versa. The smooth-
point classification problem at hand. ness term becomes a penalty if the connected points are labeled
The minimum graph cut algorithm is a powerful tool initially differently. The detailed terms in (4) and (5) are
designed for binary optimization. The classification problem
being considered here, at this stage, can be nicely regarded as a
binary discrimination problem. We categorize the whole scene (6)
content into two types which can be noted as
.
The energy function is given by (3). The first term is re-
ferred to as the data term (4) while the second term is called (7)
the smoothness term (5). and are critical weighting coef-
ficients which indicate the contributions from the two terms in where and denote two points in the point cloud and is the
the energy function. is the label that is chosen for a point. current label of point . is a threshold which is specifically
tunable. The -expansion move is used to efficiently optimize
the labeling of all points with respect to two pre-defined cate-
(3) gories and .
For the test case used, the ratio between and is chosen
(4)
to be 1:10, and the classification result is shown in Fig. 3. In
another example of result shown in Fig. 5, the top image shows
(5) a color coded DSM which clearly shows what areas are trees
with the help of visible wavelength imagery of the same study
1444 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 3, JUNE 2013
V. BUILDING DESCRIPTION
C. Terrain and Rooftop Patches Extraction
Graph cuts optimization helped to remove vegetation from A. Planar Features on Rooftop Detection
the scene. Our major working targets are man-made structures After obtaining all of the desired building rooftop patches,
(buildings specifically). So, the next step is to detect terrain and each building footprint (cluster) can be processed making use of
extract all possible rooftop patches from the scene which has some powerful computing resources built on parallel systems.
already had vegetation excluded. At this stage, it is possible to build a model from each individual
As long as various targets in the scene are spatially sepa- point set, constructing a single building from each collection.
rated from one another, it is reasonable to apply the Euclidean However, in order to achieve models containing as much fine
clustering introduced in Rusu’s work [22] to group and identify detail as possible, it is necessary to identify the many signifi-
points that make up these targets. In this approach, a strategy cant features that exist on rooftops and represent them with as
referred to as hierarchical Euclidean clustering (Fig. 4) is pro- much detail as possible. This refinement process represents an-
posed. Two consecutive clustering passes, utilizing different tol- other region growing based segmentation problem. Rabbani et
erances which determine the searching radii in the neighbor- al. [26] proposed a segmentation method using a smoothness
hood, are conducted. Generally speaking, the terrain presents constraint which avoids calculating properties like curvatures.
the largest areal coverage in many urban scenes. The first step Instead, they calculated the residual value obtained by plane fit-
attempts to separate this largest area from the rest of the scene el- ting to a small surface area and utilized it as a substitution to the
ements (the buildings). Since airborne LiDAR data is collected curvature property. Their method was only tested on surfaces
above the ground, there are very few points that fall on the sides of indoor objects in relatively small size not large surfaces like
of the buildings compared to the points that are on rooftops. rooftops. The algorithm developed in the approach presented
This fact guarantees the terrain portion of the scene will have here is very similar. The smoothness constraint is maintained,
SUN AND SALVAGGIO: AERIAL 3D BUILDING DETECTION AND MODELING FROM AIRBORNE LIDAR POINT CLOUDS 1445
(8)
Since the normals and curvatures are estimated as closely as Fig. 7. Region growing using smoothness constraint and curvature consis-
possible, yet still are represented only by their best approxima- tency.
tion, it is inevitable that over-segmentation will occur due to
the nature of the problem. A great number of tiny regions in generated. Another refinement step needs to occur to merge or
which there are only a few points, or even only one point, are combine these tiny regions with the major regions to which the
1446 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 3, JUNE 2013
Fig. 8. Each of the two example rooftops is segmented into different parts by our region growing method.
B. Boundaries Production
The next step in the presented approach is to produce pos-
sible boundary points for all features on the rooftops. These
points will also be involved in the modeling process later. The
shapes of the rooftops in our test scene have great complexities.
It is very difficult to directly fit basic geometric primitives to Fig. 9. Finding the rectilinear boundary of a group of points (blue dots). The
them. Since most common buildings have rectilinear outlines, green line indicates the outline of marked cells. The red line indicates the tight-
it is very reasonable to model boundaries of all parts of a com- ened outline after further fitting.
plex rooftop under rectilinear constraints. First, the orientation
of all features of a building rooftop is fixed to be a dominant ori-
model construction is dramatically reduced, but remains geo-
entation, which is currently determined by the prior knowledge.
metrically precise. The topological precision, however, is not
Second, a 2D grid is overlaid on the LiDAR points in the XY
guaranteed. This can lead to building rooftop outlines that ap-
plane. The dimension of the grid cell is adjustable. The cells of
pear unrealistic and distorted. In follow-on work, [16] proposed
the 2D grid are marked as being occupied if there are at least a
a topology control to be added to the 2.5D dual-contouring op-
minimum number of points in the cell. This minimum number
timization process. In this current work, every detected feature
is determined based on the density of the input point cloud.
on the rooftop is treated equally, which achieves the similar goal
The boundary of the marked cells now approximately represent
of maintaining the topology of the structures.
the shapes of all parts. The lines are adjusted to be parallel to
themselves and closer to the original points for the purpose of VI. EXPERIMENTAL RESULTS
achieving a tighter fit. The process is illustrated in Fig. 9. The proposed workflow has been tested on several airborne
LiDAR data collections carried out in the Greater Rochester,
C. Modeling
New York, USA area. The wavelength of the LiDAR is about
The 2.5D dual-contouring approach proposed by Zhou et al. 905nm, and the repetition rate is about 25 Hz. The LiDAR point
[15] is included in our approach for creating watertight building density is approximately 4 samples/m or less. Processing was
models. Zhou’s method deals with points on rooftops only and conducted on a consumer laptop (Intel Core i7, 12G RAM).
produces vertical walls connecting these rooftops to the ground. Figs. 10 and 11 show the reconstruction results for portions of
This fits the requirement of this effort. This method extends the data collected over the campus of the Rochester Institute
the classic dual contouring into a 2.5D model. It takes rooftop of Technology and the City of Rochester. The terrain, in which
points, and their estimated normals, as inputs and has no restric- there is some elevation variation, is also modeled. All models
tion on rooftop shapes. The number of points required for the are represented with small triangular facets, which is not only
SUN AND SALVAGGIO: AERIAL 3D BUILDING DETECTION AND MODELING FROM AIRBORNE LIDAR POINT CLOUDS 1447
Fig. 10. (a) Reconstruction of the buildings over the campus of RIT (terrain included). (b) The corresponding optical aerial ortho-photo (courtesy of Google Map).
valuable for visualization, but is also useful for physical simu- method which lack of the capability to capture the sides of
lation in specific applications. buildings. This means that each side of a rooftop is connected
to the ground by a simple, vertical wall which is obviously
VII. LIMITATIONS AND CONCLUSIONS not always indicative of the true architectural form. Another
There are limitations in the proposed approach. Airborne concern, which does occur frequently in residential areas, is
LiDAR point clouds are the major input to the proposed the partial occlusion of a rooftop by trees. If this occurs, then
1448 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 3, JUNE 2013
Fig. 11. (a) Reconstruction of the buildings over a part of the city of Rochester (terrain included). (b) The corresponding optical aerial ortho-photo (courtesy of
Google Map).
this portion of roof is lost in the LiDAR point cloud, which ACKNOWLEDGMENT
can result in an incomplete reconstruction. In addition, the
The authors would like to thank the United States Department
2.5D dual contouring method [15] is a robust algorithm while
of Energy for their sponsorship of this work.
not responding to our outline refinement ideally. It sometimes
introduces a certain degree of distortion to boundaries after
facetization. Some inability of the sensor itself also constrains REFERENCES
the performance of the proposed approach. For instance, some [1] J. Hu, S. You, and U. Neumann, “Approaches to large-scale urban mod-
rooftops are made of or contain areas that are constructed from eling,” IEEE Computer Graphics and Applications, vol. 23, no. 6, pp.
transparent materials, like glass. The LiDAR scan pulse typi- 62–69, Nov.–Dec. 2003.
[2] P. Musialski et al., “A survey of urban reconstruction,” in Euro-
cally does not provide a detectable return from these surfaces, graphics 2012–State of the Art Reports, 2012, pp. 1–28, Eurographics
so these regions may not be able to be reconstructed in the final Assoc..
model. [3] B. Sirmacek, H. Taubenbock, P. Reinartz, and M. Ehlers, “Performance
evaluation for 3-d city model generation of six different dsms from
In general, a method is presented that is fully automatic for air-and spaceborne sensors,” IEEE J. Sel. Topics Appl. Earth Observ.
3D building detection and modeling by processing airborne Remote Sens., vol. 5, no. 1, pp. 59–70, 2012.
LiDAR point clouds. The building footprints and the terrain are [4] T. Moons, D. Frère, J. Vandekerckhove, and L. Van Gool, “Automatic
modelling and 3D reconstruction of urban house roofs from high
first separated from vegetative areas by applying a graph cuts resolution aerial imagery,” in Computer Vision, ECCV’98, 1998, pp.
optimization technique based on the distribution of point nor- 410–425.
mals. When only rooftops and the terrain are left in the scene, [5] Z. Kim, A. Huertas, and R. Nevatia, “Automatic description of build-
ings with complex rooftops from multiple images,” in Proc. 2001 IEEE
a novel hierarchical Euclidean clustering method has been Computer Society Conf., CVPR 2001, 2001, vol. 2, pp. II-272–II-279.
developed to extract rooftop patches and the terrain. For each [6] R. Nevatia and K. Price, “Automatic and interactive modeling of build-
rooftop patch, a region growing based segmentation method ings in urban environments from aerial images,” in Proc. 2002 IEEE
Int. Conf. Image Processing, 2002, vol. 3, pp. 525–528.
has been presented that detects all significant features on the [7] J. Rau, L. Chen, and G. Wang, “An interactive scheme for building
rooftop. Boundary points are generated under some reasonable modeling using the split-merge-shape algorithm,” Int. Archives Pho-
assumptions for typical rooftop structures. Finally, the 2.5D togramm. Remote Sens., vol. 35, no. B3, pp. 584–589, 2004.
[8] N. Haala and M. Kada, “An update on automatic 3D building recon-
dual-contouring method [15] is adopted for modeling process. struction,” ISPRS J. Photogramm. Remote Sens., vol. 65, no. 6, pp.
The models utilize a UTM coordinate system and extracted 570–580, 2010.
models are ready to be embedded in to any GIS application. [9] O. Wang, S. Lodha, and D. Helmbold, “A Bayesian approach to
building footprint extraction from aerial lidar data,” in Proc. IEEE
3rd Int. Symp. 3D Data Processing, Visualization, and Transmission,
VIII. FUTURE WORK 2006, pp. 192–199.
There are some possible directions for future work. First, the [10] V. Verma, R. Kumar, and S. Hsu, “3D building detection and modeling
from aerial lidar data,” in Proc. IEEE Computer Society Conf., CVPR
lack of ground truth is a common problem in this field of re- 2006, 2006, vol. 2, pp. 2213–2220.
search, so there are no quantitative error analysis tools available [11] P. Dorninger and N. Pfeifer, “A comprehensive automated 3D ap-
at this time. The authors intend to develop a quality metric for proach for building extraction, reconstruction, and regularization from
airborne laser scanning point clouds,” Sensors, vol. 8, no. 11, pp.
these models in the near future. Second, the number of LiDAR 7323–7343, 2008.
points in a large scene is very large. Dealing with many mil- [12] C. Poullis and S. You, “Automatic reconstruction of cities from re-
lions of points efficiently is not an easy task. It is necessary to mote sensor data,” in Proc. IEEE Computer Society Conf., CVPR 2009,
2009, pp. 2775–2782.
migrate toward parallel computing techniques for this research [13] Q.-Y. Zhou and U. Neumann, “Fast and extensible building modeling
and to develop a specifically designed algorithm to process the from airborne lidar data,” in Proc. 16th ACM SIGSPATIAL Int. Conf.
data. Lastly, a method to accurately determine the principal ori- Advances in Geographic Information Systems, GIS’08, 2008, pp.
7:1–7:8.
entation of a building, or a group of buildings, in a fast, efficient [14] T. Ju, F. Losasso, S. Schaefer, and J. Warren, “Dual contouring of her-
manner is also needed. This will be a focus of future work. mite data,” ACM Trans. Graphics, vol. 21, no. 3, pp. 339–346, 2002.
SUN AND SALVAGGIO: AERIAL 3D BUILDING DETECTION AND MODELING FROM AIRBORNE LIDAR POINT CLOUDS 1449
[15] Q.-Y. Zhou and U. Neumann, “2.5D dual contouring: A robust ap- [27] M. Zuliani, “Ransac for Dummies, With Examples Using the RANSAC
proach to creating building models from aerial lidar point clouds,” in Toolbox for Matlab and More,” 2009 [Online]. Available: http://vision.
Computer Vision, ECCV 2010, 2010, pp. 115–128. ece.ucsb.edu/.../RANSAC/docs/RANSAC4Dummies.pdf
[16] Q.-Y. Zhou and U. Neumann, “2.5D building modeling with topology
control,” in Proc. 2011 IEEE Computer Society Conf., CVPR 2011,
2011, pp. 2489–2496. Shaohui Sun received the B.S. degree in biomedical
[17] F. Lafarge and C. Mallet, “Building large urban environments from engineering in 2006 and the M.S. degree in computer
unstructured point data,” in Proc. IEEE Int. Conf. Computer Vision, science and technology in 2008 from Sun Yat-sen
2011, pp. 1068–1075. University, Guangzhou, China. During his Master
[18] F. Lafarge and C. Mallet, “Creating large-scale city models from study, he did an internship at Microsoft China as
3D-point clouds: A robust approach with hybrid representation,” Int. a software development engineer. He then worked
J. Computer Vision, pp. 1–17, 2012. for Tektronix as a software engineer from 2008 to
[19] M. Ding, K. Lyngbaek, and A. Zakhor, “Automatic registration of 2010. He is currently working towards the Ph.D.
aerial imagery with untextured 3D lidar models,” in Proc. IEEE degree in imaging science at the Rochester Institute
Computer Society Conf., CVPR 2008, 2008, pp. 1–8, IEEE. of Technology, Rochester, NY, USA.
[20] L. Wang and U. Neumann, “A robust approach for automatic registra- He is a graduate research assistant in the Digital
tion of aerial images with untextured aerial lidar data,” in Proc. 2009 Imaging and Remote Sensing lab. His research interests include remote sensing,
IEEE Computer Society Conf., CVPR 2009, 2009, pp. 2623–2630. photogrammetry, computer vision focusing on three-dimensional urban recon-
[21] A. Mastin, J. Kepner, and J. Fisher, “Automatic registration of lidar and struction from airborne LiDAR and imagery.
optical images of urban scenes,” in Proc. 2009 IEEE Computer Society
Conf., CVPR 2009, 2009, pp. 2639–2646.
[22] R. B. Rusu, “Semantic 3D object maps for everyday manipulation
in human living environments,” Ph.D. thesis, Technische Universität Carl Salvaggio is a Professor of Imaging Science in
Muenchen, Munich, Germany, 2009. the Chester F. Carlson Center for Imaging Science
[23] M. Pauly, Point Primitives for Interactive Modeling and Processing of at the Rochester Institute of Technology, Rochester,
3D Geometry. Konstanz, Germany: Hartung-Gorre, 2003. NY, USA. He is a member of the Digital Imaging and
[24] Y. Boykov, O. Veksler, and R. Zabih, “Fast approximate energy min- Remote Sensing Laboratory teaching and conducting
imization via graph cuts,” IEEE Trans. Pattern Analysis and Machine research in digital image processing, remote sensing,
Intelligence, vol. 23, no. 11, pp. 1222–1239, 2001. and computer science. His particular research inter-
[25] V. Kolmogorov and R. Zabin, “What energy functions can be mini- ests lie in numerous areas including thermal infrared
mized via graph cuts?,” IEEE Trans. Pattern Analysis and Machine phenomenology, exploitation, and simulation; design
Intelligence, vol. 26, no. 2, pp. 147–159, 2004. and implementation of novel imaging and ground-
[26] T. Rabbani, F. van Den Heuvel, and G. Vosselmann, “Segmentation of based measurement systems; three-dimensional ge-
point clouds using smoothness constraint,” Int. Archives of Photogram- ometry extraction from multi-view imagery and LiDAR data; material optical
metry, Remote Sensing and Spatial Information Sciences, vol. 36, no. properties measurement and modeling; and still and motion image processing
5, pp. 248–253, 2006. for various applications.