RoboScan: an Automatic System for Accurate and Unattended 3D Scanning

M. Callieri, A. Fasano, G. Impoco, P. Cignoni, R. Scopigno

ISTI-CNR∗
G. Parrini, G. Biagini
Scienzia Machinale†

Abstract Even if the technological evolution has been impressive,

3D scanning remains a rather expensive process and under-
We describe an automatic system for fast unattended ac- utilised technology. One of the main reasons is the sub-
quisition of accurate and complete 3D models, called Ro- stantial amount of time and human intervention required.
boScan. The design goal is to reduce the three main bot- The ideal goal – making 3D scanning as easy as 2D image
tlenecks in human-assisted 3D scanning: the selection of scanning – is still far from being reached. The objective of
the range maps to be taken (view planning), the position- the research described in this paper has been to design an
ing of the scanner in the environment, and the range maps’ integrated system to make the 3D scanning of complex ob-
alignment. The system is designed around a commercial jects faster and entirely automatic. The solution proposed is
laser-based 3D scanner moved by a robotic arm. The ac- based on the adoption of commercial components: a laser-
quisition session is organised in two stages. First, an initial based 3D scanner, a robotic arm and a turntable (Fig. 1).
sampling of the surface is performed by the automatic se-
lection of a set of views. Then, some added views are auto-
matically selected, acquired and merged to the initial set in
order to fill the surface regions left unsampled. Both the ini-
tial set of range maps and the subsequently added ones are
post-processed automatically, by using the known scanner
positions to initialise the alignment phase. Results of the
assessment of the system on real acquisitions are presented
and discussed.

1 Introduction and motivation

3D scanning is recognised as the main option to re-

construct accurate and detailed models of existing objects.
Many technologies have been proposed in the last few years
to design both accurate hardware devices [5] and efficient
algorithms to post-process the sampled points [3]. Many
scanning projects concerned the acquisition of cultural her-
itage artifacts [11, 2, 19], since this domain presents inter- Figure 1. Hardware setup of the RoboScan sys-
esting case studies and general requirements (artefact shape tem.
complexity, high sampling accuracy, simultaneous sam-
pling of the reflection properties). Nevertheless, 3D scan-
ning is a general purpose technology, with a broad range of The main bottlenecks which make 3D scanning a com-
possible applications. plex task are the following:

∗ ISTI-CNR, Via Moruzzi, 1 56124 PISA, Italy; email: [email protected] • many range maps (50 ∼ 500) are produced when scan-
† Scienzia Machinale, Pisa, Italy; email: [email protected] ning a complex object; planning this set of views is

Proceedings of the 2nd International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT’04)
0-7695-2223-8/04 $ 20.00 IEEE
 not easy, since we look for the minimal set of views of the surface of scanned artifacts and to process the sam-
covering the whole object’s surface, which should sat- pled data off-line. Therefore, the attention of researchers
isfy a given minimal overlapping factor, and should be has been focused onto the issues of acquisition planning and
defined taking into account the scanner characteristics range map registration, which still remain open problems
and the orientation of the sampled surface portion; and are the crucial obstacles in the design of unattended
scanning systems. Even if a number of solutions address-
• for each view, the scanner has to be moved in the cor- ing the two problems above have been proposed, few of
responding location and orientation. Due to the weight them are feasible in practice. To our knowledge, none of
of standard commercial scanners this action is often the available scanning systems can scan accurately complex
cumbersome and time consuming; objects in a completely unattended way.
• all the acquired range maps have to be integrated in View planning. Planning algorithms can be characterised The
camera is
a single model (either triangle- or point-based); a ba- according to how much a-priori knowledge about the scene activated
continually
sic action is the mutual registration of the range maps. is available. Mainly, planners either assume that a complete
The standard solution is based on a semi-automatic model of the scene is known in advance (e.g. in the indus-
method [17], which requires an intense user contri- trial inspection of mechanical parts, where a CAD reference
bution to select the initial rough registration for each model is used to verify the compliance of manufactured ob-
scans (about 60% over the overall scanning time spent jects [23] ) or, on the other side, that no geometrical infor-
on registration). Current automatic alignment method- mation is available prior to the acquisition (as in standard
ologies (see Section 2), are not sufficiently robust. 3D scanning sessions).
In this case, the object to be scanned is assumed to lie in
Robotic arms or other type of mechanical gantries have a known bounding volume. One of the first algorithms in
been connected to scanners in the past, either with the goal this class was proposed by Connolly [6]. He partitions a
of displacing the scanner in the acquisition space or to track spherical bounding volume using an octree representation.
the scanner position [11, 9, 22]. To our knowledge, our ap- Two planning algorithms are presented. In the “planetar-
proach is innovative since it is the first case in which the ca- ium” algorithm the bounding sphere surface is sampled in
pabilities of a robotic arm are fully exploited to completely a number of candidate viewpoints. Visibility of the surface
automatise the scanning process. Its main features are: for each candidate position is evaluated and the next best
• the scanner has greater freedom thanks to the robotic view (NBV) is selected as the viewpoint maximizing the
arm and the turntable; amount of unseen nodes. The “normal” algorithm counts
the area shared by the faces of nodes which separate re-
• two basic simple solutions have been employed to ar- gions known to be empty from unseen ones. The NBV is
range an initial view planning – starting with minimal the direction maximizing the shared area. Maver and Ba-
information on the object’s shape (the bounding vol- jcsy [13] proposed a planning algorithm tailored to a light
ume extent), either a cylindrical or a planar plan can stripe range sensor constrained to translations and limited
be built. rotations in a plane above the object. Occluded regions are
represented as polygons. Viewpoint visibility constraints
• the relative positions of range maps are known, since are computed from the polygon boundaries. Whaite and
the robot movements are calibrated and accurate – Ferrie [24] developed a model-based approach in which a
ICP [4, 17] can be run without manual alignment. parametric model (superquadric) is fitted to the currently
sensed data. The shape is progressively refined minimising
• a technique is employed which determines automati- the uncertainty in the model, by scanning the region where
cally a set of views to acquire the missing range data – data fit the model worst. In Pito’s approach [15] the scan-
the coverage of the model is improved. ning volume is enclosed by a surface parameterised using
In summary, our system integrates a number of software two bi-dimensional scalar fields (positional space), encod-
techniques (some of them original) to speed-up the acquisi- ing the visibility information for each candidate viewpoint.
tion process and make unattended scanning feasible. The viewpoint which maximizes the unseen volume is cho-
sen as the NBV. Reed and Allen [18] encoded planning con-
straints as operations on sets. They consider the workspace
2 Related work as a set of 3D points. To compute viewing directions, they
apply set operators to the workspace.
The experience of people involved in some remarkable One of the most interesting algorithms for our purposes
scanning projects [11, 2, 19] has demonstrated the substan- is the one proposed by Papadopoulos-Orfanos and Schmitt
tial effort needed to obtain a sufficiently complete coverage [14]. They use a laser stripe scanner mounted on a robot

Proceedings of the 2nd International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT’04)
0-7695-2223-8/04 $ 20.00 IEEE
 with three translational DOF. Their system also incorpo- but also more complex and redundant the set of scanned
rates a path planning algorithm for collision avoidance. The samples), and the need to adopt lower quality acquisition
acquired data is used to guide both the sensor planning and techniques to sustain a real-time throughput. Moreover, if
path planning algorithms. To our knowledge this is the the ICP algorithm fails (e.g. for fast motion of the object
first attempt to assemble a fully automatic system embody- with respect to the scanner, or due to flat regions, where
ing sensor- and path-planning, together with acquisition of range maps can slide) the registration cannot be recovered.
range data. The main drawback of this approach is that the Commercial solutions are limited to the addition of a
scanning system is not allowed to rotate around the opti- tracking system to the scanner, which produces an initial
cal centre. This constrains severely the shape of scannable placement for the data acquired (solving just the registra-
objects. Moreover, their system is intended for very lim- tion problem, and not considering view planning). Stein-
ited applications since it is too tied to a specific commercial bichler’s COMET T-Scan[22] adopts optical tracking (with
hardware, handling only small objects. the inherent line-of-sight problem), while the COMET
Scott et al. [21] worked out a multi-stage approach in which VarioZoom[22] and FARO ScanArm [9] use mechanical po-
a coarse model acquired in a first stage is used to guide sitioning systems.
scene exploration for the fine modeling of the next stage.
An a-priori base model is needed at the first stage (e.g. 3 Architecture of the RoboScan system
a bounding volume). The algorithm partitions the rough
model into cavities, holes and planar patches and for each RoboScan is a fully automatic acquisition system (see
patch a set of candidate viewpoints is generated. A user de- Figure 2), composed by a commercial triangulation-based
fined measurability function is computed for each patch of 3D laser scanner (Minolta VI 910 [10]) and a commercial
the rough model. six-axis robot arm (ABB IRB 4400 [1]). To provide an-
Automatic range maps registration. There are two main other degree of freedom, a PC-controlled turntable is used
registration strategies: the adoption of a tracking system and to rotate the object in front of the robot, enabling to scan
the use of purely software solutions. the object from every direction. This heavy-duty hardware
The scanner pose can be tracked over time and information is controlled by two computational units. The unit PC1, an
about the camera transformation between scans collected. industrial all-integrated machine, is placed directly on the
Common tracking methods employ inertial, magnetic [16], robot arm (due to the limited SCSI cable length). The unit
acoustic, optical [22] or mechanical positioning systems at- PC2 is an off-the-shelf PC connected to the robot arm, to the
tached to the scanner. Tracking systems are used mainly to turntable, and to PC1. It controls the turntable and the robot
provide initial guesses to software systems, since they are arm, sends scanning commands to PC1, and post-processes
seldom sufficiently accurate. the acquired data.
Software methods (e.g. [12]) usually adopt a two-stage ap-
proach, in which a pair-wise registration is followed by a
global optimisation stage [17]. The construction of the ad-
jacency graph (i.e., find all the range maps partially over-
lapping with the one considered) and the initialisation of
the iterative ICP [4] process (i.e. find an approximate place-
ment for a pair of range maps which allows the convergence
of ICP) are the two main issues in software approaches.
Automatic scanning systems and tracked scanners.
Rusinkiewicz et. al [20] demonstrated that an automatic ac-
quisition system can be easily built using a real-time scan-
ning device. Since multiple range maps per second can be
acquired, there is a strong coherence between pairs of con- Figure 2. Hardware components and connec-
secutive range maps. Even if the object is slowly moved in tions of the RoboScan system.
front of the acquisition device, there is still a good chance of
aligning automatically the acquired samples. Furthermore,
since the model is generated on-line, the user gets some
feedback on what is going on, even if no automatic plan- 4 Unassisted 3D scanning
ning strategy has been proposed. The main drawbacks of
this approach are: the limitation on the size and weight of The strategy implemented to support unassisted 3D scan-
the scannable objects, the large number of range maps ac- ning is based on four different steps. First, an automatic
quired (the higher the frame rate, the higher the coherence, initial coverage is performed. Then, the initial set of range

Proceedings of the 2nd International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT’04)
0-7695-2223-8/04 $ 20.00 IEEE
 maps is automatically registered and merged, to produce centage are aligned with ICP. This pair-wise aligment be-
a first digital model. Successively, unsampled regions are tween overlapping range maps is then followed by a the
detected automatically and acquired. Finally, the user can global optimisation stage described in [17].
complete the model by driving manually the placement of After registration convergence, range maps are fused us-
the scanner, in case automatical detection still leaves un- ing a volumetric method based on implicit distance field
sampled regions. The integrated model can be simplified to construction and zero-distance isosurface fitting [7]. A
a user-requested size. medium resolution grid (1-2 millimetres) is used to produce
To start a scanning session, the user has to specify just a a model of adequate resolution for further processing (see
few data. The basic parameters are the target quality level Phase 3).
of the final digital model, and the extents and location of a
bounding volume enclosing the artifact. This latter data is 4.3 Phase 3: selection of missing patches
required by the system to define the initial set of scanning
poses.
The models generated so far are usually incomplete due
to self occlusions and tangential surface sections. A hole
4.1 Phase 1: rough shape acquisition detection phase is thus needed in order to automatically de-
tect unsampled regions and plan appropriate scanner poses.
At the beginning of the acquisition session there is no Different planning algorithms have been outlined in Sec-
knowledge about the shape of the object. For this reason, tion 2. Although the view planning literature is rich, none
the initial probing points have to be planned using only the of the proposed solutions is satisfactory for our needs. The
basic bounding volume data. The idea is to perform a sim- algorithms proposed in [14] and [13] depend heavily on the
ple and regular sampling, which is based on two different particular scanning configuration adopted. Solutions in [6]
scanning modes: planar and cylindrical. The planar mode are too naive and the algorithm in [24] is unable to handle
is mainly adopted when one side of the object has to be ac- complex models. Finally, as our system is designed for ap-
quired (e.g. artistic bas-reliefs or industrial workpieces). In plications working with huge meshes, key features of the
this scenario the turntable is not used. A simple grid of reg- planning algorithm should be time efficiency and low mem-
ularly spaced poses is generated, orthogonal to the side of ory occupancy. These requirements are not fulfilled by the
interest of the bounding volume (see Figure 3: the poses are approaches presented in [15, 18, 21].
represented by yellow dots, and the intermediate results are The idea underlying our planning strategy is to exploit
presented to the user interactively). The cylindrical mode the capabilities of the graphics hardware to detect quickly
is adopted for the acquisition of all-round objects, and uses and accurately unsampled regions. Then, opportune scan-
both the turntable and the robotic arm. For each rotation ner poses can be computed to acquire the missing data. The
step, a set of probing positions are taken at progressively model is rendered from a number of different candidate
increasing height, producing a sort of vertical strip of range scanner positions. Views which exhibit larger unsampled
maps shot from a direction orthogonal to the cylindrical sur- surface portions are selected as new acquisition viewpoints.
face. From each candidate viewpoint we render (disabling
The user can control the number and the density of the shading) the current model, using different colours for the
poses defined by means of two parameters: the shooting dis- background, the front-, and the back-side of its surface (Fig-
tance and the overlap among two consecutive range maps. ure 4(b)). When the back-surface is visible, we are looking
“through” missing data. Thus, back-coloured pixels repre-
4.2 Phase 2: automatic registration of range maps sent missing data. Hence, the number of such pixels is a
measure of the area of unsampled regions. It can happen
To finely register the acquired range maps we exploit the that holes on the front surface and holes on the back sur-
known scanner positional data. The robotic arm used in the face project on the same image pixels and no back-coloured
prototypal implementation has a very good positioning ac- pixel are imaged. Therefore we augment the model with a
curacy (error is  0.1mm), which in our experience is much back-coloured half-box enclosing the back of the acquired
more than the initial rough registration needed for the ICP surface. This simple trick allows to discriminate correctly
convergence. Hence, much less accurate (and cheaper) me- between exterior background and potential holes (see Fig-
chanical positioning systems might be employed as well, ures 4(a) and 4(b)). The planning algorithm starts by testing
without any concern on the quality of the 3D models pro- a set of candidate viewpoints located on a spherical surface,
duced. We build a spatial search structure (called occupancy centred in a point inside the user-defined bounding box. For
grid) to determine which range maps share approximately each rendered image, the percentage of back surface pix-
the same discretised portion of space. All the range maps els bpk is computed and the viewpoints are sorted accord-
pairs with an approximate overlaps greater than a given per- ingly. High bpk values suggest that portions of front sur-

Proceedings of the 2nd International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT’04)
0-7695-2223-8/04 $ 20.00 IEEE
 Figure 3. Automatic initial coverage. On the left, the interface used to define the bounding box and
the view planning parameters. During automatic scanning, the actual position of the robot is shown
and the object appears progressively (middle and right pictures).

face are missing in the current model. Viewpoints whose erative process. High unsampled area ratios are set in the
bpk falls below a user-defined threshold are consequently first stages to generate few good viewpoints, and then the
pruned from the set of candidate viewpoints and the remain- plan is refined lowering the area gain parameter in late it-
ing ones are sorted with respect to bpk . A viewpoint vk is erations. Another critical issue is the choice of the number
accepted if the angle between vk and v j is above a fixed of candidate views. Here the tradeoff is between a fast but
threshold for each viewpoint v j that has already been vali- sparse sampling on the one hand, and a fine but slower one
dated. View selection and testing is adaptive. We start with on the other. In our experiments, we found that a good com-
an initial sparse regular sampling of the spherical surface. promise between time efficiency and quality of plans is to
Then, directions surrounding a good viewpoint are more start with 100 candidate viewpoints, and set the area gain
densely re-sampled and tested. This process can be repeated to 0.1 and the maximum overlap to 0.3. In successive iter-
a number of times, until there is no appreciable change in ations the area gain is halved at each step, while the other
the most significant section of the sorted view list, or up to parameters remain unchanged.
a maximum level of recursion.
The snapshot presented in Figure 5 shows the acquisition
This simple algorithm can be easily extended to handle
of the range map corresponding to the pose shown in Fig-
two-line-of-sight scanners. For each pose vi selected by the
ure 4(b).
procedure described above, we search for an unoccluded
line-of-sight for the laser beam. Taking into account the 3D
Our planning algorithm meets the initial requirements.
scanner geometry, new views are rendered around vi and the
Memory consumption is kept low, since we use a low res-
same thresholding operation is applied to the new frames. If
olution representation of the mesh and a copy of the frame
no good view is found then we assume that no good scanner
buffer. Processing time is low as well, since we take advan-
position exists from that view, and we discard vi . Otherwise
tage of modern GPU rendering speed.
we select the best view as the position of the laser beam.
The planning algorithm can be tuned by setting a few pa-
rameters: the minimum unsampled area ratio (area gain),
the maximum overlap between views, and the number of As the scanning process goes on, more and more empty
candidate viewpoints we are searching in a single iteration. volume is detected in the initial bounding volume. It can
In particular, a low value of area gain allows the planner to be safely be “carved out” of the current bounding volume.
detect even tiny holes, but tends to generate too many valid This allows a more accurate selection of the range maps
viewpoints. Even if this can be partly handled by tuning the to be acquired. Collisions are prevented, since a surface
maximum-overlap parameter the problem still holds. That enclosing the object is always known to the path planning
is why we have designed the view planning stage as an it- unit of the robotic arm.

Proceedings of the 2nd International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT’04)
0-7695-2223-8/04 $ 20.00 IEEE
 (a) The same model is shown with (left) and without (right) shad-
ing. The yellow frame shows the region sampled by scanner from
a given viewpoint.
Figure 5. Automatic selection and acquisition
of missing range maps: before and after the
inclusion of a new range map (in yellow) from
a new viewing direction (red line).

user has to select a target point on the digital model (or

on the bounding volume) and to specify which incidence
(b) A pink half box is added to detect all missing surface portions. angle should be used. The computation of the safe path
which moves the scanner head from the rest position to the
Figure 4. The front and the back side of the selected pose is computed by the robot driver, taking into
surface are rendered in blue and pink, respec- account the geometry of the scene to prevent collisions.
tively. The background is rendered in black.
The ratio of pink pixels in the image is a mea-
sure of the area of unsampled regions. 5 Results

Tracking accuracy. The positional information tracked by

our system is totally adequate to our purposes. The robot
has a declared positioning precision of  0.1mm [1] which
4.4 Phase 4: surface completion and generation was confirmed in our assessment. In a number of appli-
of the final model cations, this level of accuracy makes the further software-
based registration of the range maps an optional step (or,
When the resulting set of range maps is considered suf- on the other hand, opens room for the experimentation of
ficiently complete, the final mesh post-processing actions lower accuracy and cheaper robotic arms). Conversely, the
are started. The alignment of all range maps acquired is low-cost turntable used has proven to be less accurate (for
further refined. Then, range maps are merged in a trian- mechanical imprecision and small errors during axis cali-
gle mesh using a volumetric reconstruction approach. The bration). The positional data returned are accurate enough
resolution depends on the final accuracy requested 1 . The to initialise the ICP process, but, without a ICP pass they
small holes which could still be present in the surface, can are not sufficient for a high quality reconstruction.
be closed adopting an approach based on volumetric diffu- Acquisition speed. The most important results is, in our
sion [8]. The last step performs some controlled simplifica- opinion, the speed of the system. The speed-up with respect
tion of the triangle mesh produced, to reduce the unneces- to manual acquisition reaches at least one order of mag-
sary data fragmentation introduced by the Marching Cubes nitude. In various tests the system has proven to be able
fitting. The final model can be exported in various file for- to capture 30 range maps (i.e. the number of range maps
mats (PLY, VRML, STL, Point Cloud). needed to cover one square metre of object’s surface) in 20
An interface to drive the scanner in an assisted way is minutes. Alignment time is completely cut out in the planar
provided to accomodate for very complex objects. Since set-up, and reduced to 3-4 minutes every 30 range maps in
moving a six-axis robot arm directly is not an easy task, the cylindrical set-up. Times for the generation of the final
the software helps the user, simplifying the process. The model (reconstruction and simplification) cannot be given
1 1mm
as a single number since they depend on the model size, the
is the standard value necessary for milling-based reproduction,
number of range maps, and the resolution requested. The
0.2-0.4mm are used for rendering and archival purposes or in the case
of rapid reproduction with more sophisticated devices, such as stereo- 3D model of the Pharaoh bust shown in Figure 6 (height
lithography 50cm, width 30cm) has been produced in about 30 minutes,

Proceedings of the 2nd International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT’04)
0-7695-2223-8/04 $ 20.00 IEEE
 Acknowledgements We acknowledge the financial support
of the projects EU IST-2001-32641 “ViHAP3D” and Re-
gione Toscana “DigiSculptor”.

References

[1] ABB and Ltd. Articulatet arm robot irb 4400.

http://www.abb.com/robots, 2004.
[2] F. Bernardini, J. Mittleman, H. Rushmeier, and G. Taubin.
Case study: Scanning Michelangelo’s Florentine Pieta’. In
ACM SIGGRAPH 99 Course Notes, Course 8, August 1999.
[3] F. Bernardini and H. E. Rushmeier. 3D Model Acquisition.
In Eurographics 2000, State of the Art Reports Proceedings,
Figure 6. Two test objects: a marble capitello pages 41–62. Eurographics Association, Aug. 24–25 2000.
[4] P. J. Besl and N. D. McKay. A method for registration of 3-D
(50cm tall 70cm wide) and an Pharaoh bust.
shapes. IEEE Transactions on Pattern Analysis and machine
Intelligence, 14(2):239–258, Feb. 1992.
[5] F. Chen, G. Brown, and M. Song. Overview of three-
including: system startup, first acquisition round (about 30 dimensional shape measurement using optical methods. Op-
tical Engineering, 39(1):10–22, January 2000.
range maps), alignment, planning and acquisition of addi- [6] C. Connolly. The determination of next best views. In
tional range maps, and the final reconstruction at 0.5 mm. CRA85, pages 432–435, 1985.
For comparison purposes, a recent manual acquisition of a [7] B. Curless and M. Levoy. A volumetric method for building
marble statue (height 80 cm, width 35cm) was performed complex models from range images. In Comp. Graph. Proc.,
using standard post-processing tools, not tracking the scan- Annual Conf. Series (SIGGRAPH 96), pages 303–312. ACM
ner motion, with the statues placed on a turntable. We spent Press, 1996.
[8] J. Davis, S. Marshner, M. Garr, and M. Levoy. Filling holes
6 hours to acquire the data (around 100 range maps) and
in complex surfaces using volumetric diffusion. In First Int.
18 hours to align them manually. For a completely manual Symp. on 3D Data Processing, Visualization and Transmis-
acquisition (without the turntable) we estimate 12 hours for sion (3DPVT’02), pages 428–438. IEEE Comp. Soc., 2002.
scanning and a couple of days for alignment. [9] FARO Technologies Inc. ScanArm.
In order to show the accuracy of our system, no hole filling http://www.faro.com/, 2004.
has been run on the models shown in Figures 6 and 7 [10] Konica Minolta. Minolta VI 900/910 high resolution laser
scanners. http://www.minolta-3d.com/, 2004.
[11] M. Levoy, K. Pulli, B. Curless, S. Rusinkiewicz, D. Koller,
6 Conclusions L. Pereira, M. Ginzton, S. Anderson, J. Davis, J. Gins-
berg, J. Shade, and D. Fulk. The Digital Michelangelo
We have presented the architecture of a system for unat- Project: 3D scanning of large statues. In SIGGRAPH 2000,
tended 3D scanning of complex artifacts. The system Computer Graphics Proceedings, Annual Conference Series,
adopts a robotic arm and a turntable to move the scanner pages 131–144. Addison Wesley, July 24-28 2000.
[12] L. Lucchese and G. M. Cortelazzo. A noise-robust frequency
on any selected view, under computer control. The system domain technique for estimating planar roto-translations.
includes all the usual features to process range maps (align- IEEE Trans. on Signal Processing, 48(6):1769–1786, 2000.
ment, reconstruction, simplification) together with innova- [13] J. Maver and R. Bajcsy. Occlusions as a guide for planning
tive solutions to select and register in an automatically the the next view. PAMI, 15(5):417–433, May 1993.
set of range maps covering the object [14] A. Papadopoulos-Orfanos and F. Schmitt. Automatic 3d dig-
We tested the system on many sample objects. We ob- itization using a laser rangefinder with a small field of view.
In 3DIM97, pages 3 – View Planning, 1997.
tained very good results, both in terms of completeness of [15] R. Pito. A solution to the next best view problem for auto-
the digital models produced, and speed of acquisition. The mated surface acquisition. PAMI, 21(10):1016–1030, Octo-
first significant commercial application of this system was ber 1999.
the acquisition of the Deposed Christ shown in Figure 7. [16] Polhemus. The FastSCAN Co-
It is a work of art sculpted on light material (polystyrene) bra scanning system. More info
whose reproduction on marble was commissioned to Scien- on:http://www.polhemus.com/fastscan.htm,
zia Machinale srl. The digital model obtained has been eval- 2003.
[17] K. Pulli. Multiview registration for large datasets. In Proc
uated very positively by the reproduction experts and by the 2nd Int.l Conf. on 3D Digital Imaging and Modeling, pages
customer. The artifact is now being reconstructed on marble 160–168. IEEE, 1999.
from the digital model. The milling system used employs [18] M. Reed and P. Allen. Constraint-based sensor planning for
the same robotic arm. scene modeling. PAMI, 22(12):1460–1467, December 2000.

Proceedings of the 2nd International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT’04)
0-7695-2223-8/04 $ 20.00 IEEE
 Figure 7. The RoboScan system acquiring the “Deposed Christ" high-relief (real size: 2.5 × 1.5 metres),
and the corresponding digital model produced.

[19] C. Rocchini, P. Cignoni, C. Montani, P. Pingi, R. Scopigno,

R. Fontana, L. Pezzati, M. Cygielman, R. Giachetti, and
G. Gori. 3D scanning the Minerva of Arezzo. In ICHIM’2001
Conf. Proc., Vol.2, pages 265–272. Politecnico di Milano,
2001.
[20] S. Rusinkiewicz, O. Hall-Holt, and M. Levoy. Real-time 3d
model acquisition. In Comp. Graph. Proc., Annual Conf. Se-
ries (SIGGRAPH 02), pages ??–?? ACM Press, July 22-26
2002.
[21] W. Scott, G. Roth, and J.-F. Rivest. Performance-oriented
view planning for automatic model acquisition. In Proceed-
ings of the 31st International Symposium on Robotics, pages
314–319, 2000.
[22] Steinbichler Optotechnik. COMET series.
http://www.steinbichler.de/, 2004.
[23] G. Tarbox and S. Gottschlich. Planning for complete sensor
coverage in inspection. CVIU, 61(1):84–111, January 1995.
[24] P. Whaite and F. Ferrie. Autonomous exploration: Driven by
uncertainty. In McGill, 1994.

Proceedings of the 2nd International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT’04)
0-7695-2223-8/04 $ 20.00 IEEE