The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVIII-2/W4-2024
10th Intl. Workshop 3D-ARCH “3D Virtual Reconstruction and Visualization of Complex Architectures”, 21–23 February 2024, Siena, Italy
ENHANCING TERRESTRIAL POINT CLOUDS USING UPSAMPLING STRATEGY:
FIRST OBSERVATION AND TEST ON FARO FLASH TECHNOLOGY
Giulia Sammartano1,2, Giacomo Patrucco1,2, Marco Avena1,2, Cristina Bonfanti3, Antonia Spanò1,2
1
DAD - Department of Architecture and Design / Politecnico di Torino, Viale Mattioli 39, 10125, Torino (Italy).
giacomo.patrucco@polito.it; giulia.sammartano@polito.it; marco.avena@polito.it, antonia.spano@polito.it
2 FULL – Future Urban Legacy Lab, Politecnico di Torino. Via Agostino da Montefeltro 2, 10125 Torino (Italy).
3 CAM2 (FARO Technologies, Inc.) AEC Team Italy
KEYWORDS: flash technology, upsampling algorithms, TLS, point clouds density, quality evaluation, FARO Technologies.
ABSTRACT:
Nowadays 3D digitization through the combination or hybridization of different sensors, with the final aim of accelerating the phases
of data acquisition and storage, develops user friendly and robotics systems, making efficient the operator role. New technologies as
Hybrid Reality Capture™ (HRC), with Flash Technology (FARO Tech.) certainly fits into this market trend, and it is characterized by
rapid acquisition, involving 3D scanning data with panoramic images contribution. The system is still under patent, and nothing is yet
released on the technology. This research presents the analysis and discussion of results based on the raw and processed data related
to the new FARO system. The assumption – based on the information declared by the manufacturer (FARO, 2023) – is that the new
colored Flash scans are faster and denser than scans of the same resolution obtained using traditional static scanning method, due to
the crucial contribution of the PanoCam data and resolution on which the upsampling strategy is based. An evaluation based on detailed
analysis of the upsampling results is reported, delivering that the surface point density exponentially decreases with the distance and
with the ray incidence inclination. A comparison with a mobile mapping technology is finally presented and discussed.
1. INTRODUCTION
For many years, research and consequently the measurement
systems market make users familiar with the continuous
development and offers of increasingly automated solutions,
which optimize 3D digitization through the combination or
hybridization of different sensors, with the final aim of
accelerating the phases of data acquisition and storage, making
user friendly the use of the system, when these solutions do not
seek to replace the operator via robotic systems.
According to this trend, FARO Technology is developing and
presenting to the user community a new solution combining the
series S Faro Scanners with the new Hybrid reality capture
(HRC) release coupled with a panoramic camera: this solution
allows acquisitions of ultra-fast point clouds, with a lower density
than the corresponding classic clouds from static acquisitions,
and produces a final cloud comparable in density to the standard
one, exploiting an upsampling technique called flash technology
connected to the simultaneous pano-camera image acquisition.
Although FARO technology is still under patent, and nothing has
yet been released on the technology, the research group has had
the opportunity to observe and assess these data with the purpose
of highlighting limitations and potential in the heritage sector. In
fact, the system declares its limits of applicability for large
project with wide areas, bottlenecks, great details, high fidelity,
as well as potential. The aim of this research tries to verify and
validate them in the heritage building complexes context.
The preliminary investigation wants to assess the new FARO
HRC Flash Technology and the system performance in an
indoor-outdoor heritage complex scenario. As a first step, the
research focuses on the porch and courtyard area of the Royal
Palace in Turin (Figure 1).
1.1 Upsampling strategies: promising perspective
The problem of increasing image resolution and point cloud
resolution and density is crucial in many application sectors;
Traditionally, in the mapping realm, the need to apply upscaling
techniques has been much investigated to address the problem of
spatial resolution of satellite images for their use with remote
sensing techniques. (Riihimäki et al., 2019, Ajmar et al. 2017)
Furthermore, in the field of security and surveillance from imagebased systems, the problem of deriving high-resolution (HR)
images from low-resolution images (LR) exploiting SuperResolution (SR) models has undergone an extraordinary
development, also in the direction of strategies for recognizing
people (person Re_Identification) from surveillance cameras.
(Hauptmann et al. 2016)
In this framework, many advances were able to benefit from
advanced network structure and deep learning strategies (Zhang
et al. 2021; Li et al. 2019, Charles et al. 2017), and also many
studies address the problem of different image resolution by
employing cross-resolution approaches. (Jiao et al. 2018).
Taking a step back, it can be said that the point cloud upsampling
algorithms took advantage of previous research on those intended
for image upscaling, which is a typical computer vision problem,
and a general classification of solutions exploiting different
approaches can be as follows:
⁻ Deep learning-based super-resolution methods (Kim & Lee,
2018);
⁻ Random forests recognised as highly non-linear learners that
can handle high dimensional noisy inputs (Schulter et al. 2015).
One of the substantial differences that exist between the problem
of increasing the resolution of single images and point clouds is
that the 3D cloud is very different from a 2D grid since clouds
derived both from Lidar or image-based techniques always
present scattered points with the non-regular spatial distribution.
If the criticalities of upscaling in the field of images can be linked
to camera settings, points of view, lighting problems and
background changes, the point clouds upsampling strategies must
face both similar problems possibly related to the sensors, as well
as any possible moving objects, the shooting distances and most
of all problems connected to the morphology of the scene that has
a great influence on the cloud noise which is one of the most
relevant challenging issues.
Furthermore, the cloud generated by the upsampling technique
must necessarily continue to effectively describe the reference
surface (object) represented by the cloud. Therefore, the issue is
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLVIII-2-W4-2024-381-2024 | © Author(s) 2024. CC BY 4.0 License.
381
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVIII-2/W4-2024
10th Intl. Workshop 3D-ARCH “3D Virtual Reconstruction and Visualization of Complex Architectures”, 21–23 February 2024, Siena, Italy
surely more complex than a simple interpolation, which is
definitely not sufficient (Yu et al. 2018).
As will be better deepened in the following paragraphs, if cloud
densification is a widely addressed problem, the most
challenging issue is, above all, when densification involves an
increase in the level of detail of the cloud, and therefore also a
problem of defining the edges of the surfaces.
However, a very important issue to consider is that the raw point
cloud data significantly benefit from upsampling techniques,
thanks to the noise decrease and the greater uniformity of the
cloud obtained, which means the more regular spatial distribution
of points enables or helps further processing that the cloud may
be subjected to.
In fact, the use of such optimized point clouds is required or
welcome in many different directions such as classification,
detection and segmentation of 3D surfaces. (Zhang et al. 2022)
Precisely Zhang et al. (2022) allow a clear and critical framework
of the many different attempts to solve the upsampling problem,
classifying in their review the methods developed so far into
optimization-based point cloud upsampling and deep learningbased point cloud upsampling, the latter both supervised and
unsupervised (Figure 1).
Moreover, other starting points from strategies aimed at the same
objectives and applied in neighboring fields, as well as already
very settled, come from (Park et al. 2011) who studied the
possibility of improving the resolution of point clouds acquired
by TOF 3D cameras, notoriously characterized by low resolution,
improving the edges of objects in depth maps using high
resolutions RGB inputs.
2. METHODOLOGY
The research aim is to evaluate the performance of the Flash
Technology (FARO Tech.) system – based on the collection of
short and low-resolution static scans processed with an
upsampling algorithm – in terms of acquisition efficiency and
delivered final data. The assumption – based on the information
declared by the manufacturer (FARO, 2023) – is that the new
colored Flash scans (Figure 2) are faster and denser than scans of
the same resolution obtained using the traditional static scanning
method, due to the crucial contribution of the PanoCam data and
resolution – equipped on the system – on which the upsampling
strategy is based. This feature represents a groundbreaking
advancement since it allows to acquire high-resolution data –
characterised by geometric definition and level of detail
comparable to traditional data – with a significant time saving
from an acquisition and processing perspective.
2.1 The scanning system method
The Hybrid Reality Capture™ (HRC), with Flash Technology
(FARO Tech.), certainly fits into the introduced market trend of
sensors hybridization and data acquisition phase acceleration.
Compared to the previous data collection technology, the Flash
Technology has been implemented as an optimization of both
acquisition rapidity and data precision and quality (FARO,
2023).
Figure 1. Synthesis of the classigfication of upsampling
approaches (Zhang et al. 2022)
Referring to the diagram in Figure 1, which we produced on the
basis of Zhang's reflections, regarding the left column, since
optimization-based methods are not based on a data-driven
approach, they require a priori data such as the evaluation of the
normals and they rely on the regularity of the object surface, so
they present a number of limitations. This is one of the most
relevant reasons for developing alternative and most effective
solutions, such as deep learning-based methods (right both
columns). Among supervised and unsupervised methods, the first
category relies on network training learning from the
downsampling process, while unsupervised upsampling
solutions don’t need priors downsampling manually conducted,
so in this perspective, they are preferred. (Zhang et al. 2022).
Lastly, we would like to cite a couple of strategies that propose
workflows combining point cloud upsampling in combination
with image-based integrations or fusion-based solutions, since
the new solution by Faro exploits the pano camera images to
reach the results.
The first one, from Nguyen et al. 2022, suggests combining raw
point clouds with being upsampled in combination with 2D
images from which they extract more information with the use of
a generative Adversarial Network (GAN) in the training and
testing phase. This method has been applied to the detection and
segmentation of cracks patterns, for example, pertaining to
bridges, buildings, or other infrastructures coming from the
construction sector, so very close and relevant to the domain the
present paper focuses.
(a)
(b)
Figure 2. (a) Preview of a collected Flash scan; (b) Spherical
image acquired with Panocam.
The so-called Flash scanner system (Figure 2) is characterized by
a speed of 10 seconds and coupled with a panoramic camera (less
than 30 seconds of data capturing, considering both range and
image data), direct traditional TLS technologies towards the
speed of mobile systems (MMS) based on portable scanners. This
has already been faced by (Bonfanti et al. 2021) for the FARO
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLVIII-2-W4-2024-381-2024 | © Author(s) 2024. CC BY 4.0 License.
382
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVIII-2/W4-2024
10th Intl. Workshop 3D-ARCH “3D Virtual Reconstruction and Visualization of Complex Architectures”, 21–23 February 2024, Siena, Italy
Swift scanner, as an evolution of the traditional terrestrial laser
scanning approach. New Flash tech. is undoubtedly interesting
because it combines rapid acquisition, involving horizontal
scanning angular step of 10'' with the panoramic images
contribution. This allows the implementation of an upsampling
strategy to the point cloud which the present research tries to
investigate and describe.
Specifically, during the data acquisition, the scanning plan is
managed by the operator through a mobile device (e.g., tablet)
where it is possible to assist and operate, if necessary, a semiautomatic visual pre-alignment based on the scan’s positions and
scanner movements.
This introduces an improvement also for the time-consuming
subsequent registration phase. In the processing phase,
subsequently, the upsampling algorithm (under FARO patent) is
the solution for the point cloud density and colour content. Here
a crucial role is played by the PanoCam, integrated into the
scanner and whose centers have been calibrated in order to
associate radiometric content to the scan data and mainly, as
declared, to improve point density: the camera is the Ricoh Theta
Z1 360° (7296*3648px resolution) (Figure 2b).
The capturing technology is based on an extremely faster
acquisition phase, as introduced, and a hybrid scan processing
exploiting both the low-resolution raw scans and the highresolution PanoCam images contribution (Figure 2b). The
acquisition parameters for the Flash scans, according to the
consolidated FARO settings, are (2x) quality and (¼) resolution,
that correspond to declared point spacing 6mm@10m.
a first-step point cloud assessment. In these directions, many
considerations can be addressed in order to understand the results
of the upsampling algorithm.
Based on the conducted analysis, one of the most important
aspects affecting the upsampling performance is the scan
acquisition configuration: scan position, distances and rays’
directions, according to the digitized surfaces, as clearly visible
in Figure 5. Actually, the effectiveness of the upsampling
algorithm implemented with the Flash tech. depends on the
relationship between the angular resolution (par. 3.1) and the
surface orientation with respect to the angle of incidence
dimensions (par. 3.4) and, of course, of the detail to be detected.
(a)
2.2 Flash data processing
The scans project is based on a first calibration scan in higher
quality, ensuring the accuracy of hybrid LiDAR-panoramic data
acquisition. In the registration phases, it is allowed to exploit
ICP-based and target-based approaches too for accuracy control.
The point cloud preview is visible in Figure 3. The overall
statistics are: ICP-quality = 1.5 mm (83% points deviation
<4mm) and Markers-quality mean error=11 mm, (st.Dev=4mm).
In Figure 4 an example of a orthophoto valorising the high quality
radiometric data.
Figure 3. Flash point cloud of the Royal Palace’s cloister.
(b)
(c)
Figure 5. The courtyard façade in different visualization: (a) the
scan planar preview, (b) the farther surface from the scan
position, and (c) the different upsampling performance on
foreshortened (façade) and reduced inclination (arches).
3.1 Scan pattern density: expected VS measured
Figure 4. Orthophoto derived from Flash point cloud.
3. RESULTS AND ANALYSIS
The results analysis and discussions will be oriented toward
different points of view and based on the output data analysis and
The first aspect to be considered in detail is the characteristic of
pattern for the final Flash point cloud. The test will compare the
declared pattern density with the output data resolution and size.
A single raw Flash scan consists of almost 25mln points and
500Mb, compared to a generic static one, of 45mln points per 700
Mb. For a Flash scan, the declared planned pattern density is
1918*4267 points (for a ¼ res.). This, although, corresponds to a
final scan resolution of about 8000*3300 points per scan
(according to the angular step, almost 0°3’, Par. 3.2).
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLVIII-2-W4-2024-381-2024 | © Author(s) 2024. CC BY 4.0 License.
383
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVIII-2/W4-2024
10th Intl. Workshop 3D-ARCH “3D Virtual Reconstruction and Visualization of Complex Architectures”, 21–23 February 2024, Siena, Italy
This final resolution is thus comparable to an approximately
25mln points scan analysed in this project, and not 8mln points
expected from the pattern density (Table 1).
From scanner Spec
After Upsampling
(metadata)
(export data)
1918*4267
8000*3300
~ 8 mln points
~ 20-25 mln points
Table 1. Table with the comparison between the resolution
parameters as reported in the scanner metadata and the
resolution of the point cloud after the upsampling.
3.2 Point cloud angular spacing
Considering the declared angular measurement, for traditional
and Flash scans, the density analysis has been thus conducted
firstly for angular spacing (Par. 3.2) and then for surface density
(Par. 3.3.).
If the starting scans parameters are equivalent (¼ resolution, 2x
quality), a standard FARO Focus scan reports a point density of
6mm@10m (These values are declared by manufacturer.
However, there are no indications regarding the same density
values of the Flash technology scans). In fact, in the Flash the
empirically verified data is almost 8mm@10m (Figure 6) an in
Table 2 a systematic analysis of angular spacing at different
distances.
The first comparison is between a couple of scans acquired in the
middle of the cloister, where the measured surfaces are located at
the greatest distance from the sensor during the scanning process.
A density analysis has been carried out using the number of
neighbour method (for each point, a sphere of 0.03 m has been
computed, evaluating the number of points inside the sphere). As
expected, the number of points in the traditional scan is
significantly higher – approximately double – compared to that
of the Flash scan (traditional scan: approximately 27 mln of
points; flash scan: approximately 15 mln of points) and
consequently, the density of the two scans has a ratio of
approximately 1:2, as observed in Figure 7. In both scans, the
density variation exhibits a radial behaviour and is mainly
influenced by distance: the closer areas – specifically, the surface
of the flooring to the instrument during the acquisition – are
denser both in the static scan and the Flash scan, as expected.
Additionally, regarding the more distant surfaces (in this case,
>30 m), despite the lower density the traditional scan manages to
capture some architectural details (e.g., mouldings, window
elements, etc.) whereas the Flash point cloud is sparser and a
higher number of gaps can be observed, especially where the
incident angle of the laser beam is greater (as it is possible to
observe in Figure 8). This indicates that the employed
upsampling algorithm is less effective at medium to long
distances.
Angular
Surfaces points
spacing (°)
spacing (mm)
1.5m
~ 0°3’
~1.5mm
5m
~ 0°3’
~5mm
10m
~ 0°3’
~8mm
10m (foreshortened)
~ 0°3’
~15mm
50m
~ 0°3’
~40mm
70m (foreshortened)
~ 0°3’
~260mm
Table 2. Angular spacing values at fixed distances in terms of
angles (°) and point spacing (mm).
Distance
Figure 6. The setup of the 3D points spacing analysis
3.3 Point cloud density
A significant aspect that considerably affects the densification
process derived from the upsampling algorithms implemented in
the analysed scanning system is represented by the acquisition
distance. This is evident by observing the comparison between
the density analyses carried out on point clouds acquired with the
laser scanner placed at different distances from the analysed
surface, considering both the traditional static method – as the
ground truth model – and Flash acquisition.
Figure 7. Density analyses carried out on the scans acquired in
the center of the cloister: (a) Flash scan; (b) Traditional static
scan.
A second test has been carried out with the aim of evaluating the
upsampling behaviour on surfaces located at shorter distances.
Therefore, two scans acquired under the vaults of the porch have
been considered for this analysis. Again, it emerges that the Flash
point cloud is composed of a significantly lower number of points
(traditional scan: approximately 36 mln of points; flash scan:
approximately 21 mln of points) but in this case the area where a
higher density of points is observed belongs to the Flash scan. In
fact, as it is possible to observe in Figure 9a and b, in the
proximity of the laser scanner position, the density between static
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLVIII-2-W4-2024-381-2024 | © Author(s) 2024. CC BY 4.0 License.
384
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVIII-2/W4-2024
10th Intl. Workshop 3D-ARCH “3D Virtual Reconstruction and Visualization of Complex Architectures”, 21–23 February 2024, Siena, Italy
scan and Flash scan is comparable. It is also important to
emphasise that in the area where the higher density is observed
(the surface of the vault above the laser scanner during the
acquisition), the Flash scan surpasses the traditional scan with a
ratio of approximately 3:1, noticing in the higher point a density
of more than 20,000 pts/V sphere r = 0.03 m. This phenomenon
demonstrates, as stated by the manufacturer, that this technology
– and consequently, the implemented upsampling algorithm –
performs better and allows for obtaining denser and
geometrically defined results at short distances, while this
technology becomes less effective in the densification task as the
distance and the angle of incidence of the laser beam increase.
3.4 Point cloud resolution performance
It has been analysed that the upsampling performance is better where
no strong edges are detected on the surveyed surface. In the case of
corners, as expected, the reconstructive behaviour struggles to
generate dense detail with respect to the plane. Furthermore, also the
shape and extension of the geometry to be measured also make a
difference in the details result: since the vertical spacing is different
from the horizontal, there are different behaviours in relation to
vertically-developed objects compared to horizontally-extended
ones.
In fact, the Flash scan project has been organized and performed
according to a global uniformity of coverage and density, taking into
consideration scans position, distances and rays’ inclination and to
study their influence on the upsampling algorithm.
It is now interesting to evaluate how the surface density is influenced
by the ray incidence on objects. For this reason, the following
analysis has been carried out with the aim of stressing and evaluating
the variation – in terms of performance – of the upsampling
algorithm as the incident angle of the laser beam changes with
respect to the detected surfaces. Different Flash scans have been
considered and the acquired surfaces have been analysed in terms of
completeness, focusing on the facades of the porch. This is due to the
fact that, as stated by the manufacturer (which declares that this
strategy is particularly suitable for surveys carried out at short range),
the most challenging surfaces to reconstruct are those located at a
greater distance from the sensor (45-65m). In Figure 10 (scan022, in
the upper porch corner), Figure 11 (scan015 in the right part of the
porch) and Figure 12 (scan010 acquired in the middle) it is possible
to observe how the acquisition pattern, result of the upsampling
patented algorithm, varies – in terms of geometric completeness –
with the movement of the relative position of the laser scanner.
Figure 8. (a) Position of analysed surfaces (b) Detail of the
density analyses carried out on the Flash scan acquired in the
center of the cloister; (c) the density analyses carried out on the
traditional static scan acquired in the center of the cloister.
(a)
(a)
(b)
Figure 9. Density distribution analyses carried out on the scans
acquired in the porch: (a) static one, under the vaults of the
cloister, (b) flash scan, under the porch arch.
(b)
Figure 10. (a) Flash scan 3D view; (b) Analysis carried out on
the scan acquired in the 022 position.
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLVIII-2-W4-2024-381-2024 | © Author(s) 2024. CC BY 4.0 License.
385
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVIII-2/W4-2024
10th Intl. Workshop 3D-ARCH “3D Virtual Reconstruction and Visualization of Complex Architectures”, 21–23 February 2024, Siena, Italy
(a)
(b)
Figure 11. (a) Flash scan 3D view; (b) Analysis carried out on
the scan acquired in the 015 position.
(a)
(b)
Figure 12. (a) Flash scan 3D view; (b) Analysis carried out on
the scan acquired in the 010 position.
Specifically, the areas where the point density is higher and the
geometric definition of the façade is comparable to a traditional
TLS point cloud are evidenced in green. Areas evidenced in
yellow are those where a significantly lower density is observed
despite the acquisition distance being comparable or equal to the
neighbouring surfaces (Areas in red are expected occlusions
generated by the pillars of the porch, for the single scan position).
From a visual inspection, it is immediately evident that there is a
significant correlation between the sensor position and the
orientation of the measured surface, particularly in terms of
façade orientation with respect to the incident laser beam.
When the orientation of the façade is perpendicular (frontal
position) to the sensor signal, the most complete results – in terms
of geometric reconstruction – are observed in a specific angular
range (approximately) between 60° and 90°, verified in different
cases. As the inclination of the surface increases (becoming
evident on the facades located laterally to the laser scanner
position), the data becomes sparser and scarcer. In this case, it is
observed that the best reconstruction performs when the angular
value between the incident laser beam and the analyses surfaces
falls within a range higher than 30°-40°.
This aspect becomes further evident from the local analyses
carried out on different samples (dimension of samples: 1m x 1m)
with the aim of evaluating the algorithm performance at fixed
distances. In fact, the surface point density exponentially
decreases with the distance and with the ray incidence
inclination.
Two ranges have been considered: 10m (short) and 50m (medium
range) (Figure 13a). Regarding the density values – observed in
samples analysed at 10m – are the following:
• 16954 points/m2 @10m (orthogonal direction of the laser beam
on the analysed sample);
• 9206 points/m2 @10m, (foreshortened direction of the laser
beam on the analysed sample).
As evident from these values, the orientation of the object
acquired by the laser beam heavily affects the density of the final
point cloud. In this case, foreshortened objects are characterised
by a density which exhibits a ratio of 1 to 2 in comparison to
orthogonal surfaces.
This is even more evident when observing the distribution of the
points extracted from the samples acquired at 50 m (in this case,
a semi-circular niche containing a statue was considered, Figure
13a). In Figure 13b, when the surface is approximately
perpendicular (eg., the wall adjacent to the niche), the density of
the points and the quality of the geometric reconstruction are
comparable to the sample extracted in the same position from a
traditional TLS point cloud (in both cases the observed point
spacing is approximately 35-40 mm).
However, in Figure 13c it is visible that the upsampling algorithm
completely failed to reconstruct the inclined and curved areas of
the niche, with point spacing > 40 cm. However, the only area of
the semi-circular niche where the surfaces have been properly
measured is the central area with the incident angle is almost 90°.
(a)
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLVIII-2-W4-2024-381-2024 | © Author(s) 2024. CC BY 4.0 License.
386
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVIII-2/W4-2024
10th Intl. Workshop 3D-ARCH “3D Virtual Reconstruction and Visualization of Complex Architectures”, 21–23 February 2024, Siena, Italy
•
•
(b)
•
•
(c)
Figure 13. The porch apse, (a), visualized on a scan in the
opposite (b) FARO static, (c) FARO Flash.
4. DISCUSSION: A NEW MMS SOLUTION?
One of the undeniable advantages associated with the use of this
FARO Flash method concerns the optimization of efficiency
compared to the traditional TLS approach. In connection with
this concept of acquisition rapidity, the recent trends in the MMS
domain should be considered in this case, discussing this new
technology. Regarding rapid mapping strategies aimed at
architectural 3D sensing and documentation, a predominant role
is played by the use of mobile SLAM-based solutions, which are
capable of ensuring performance efficiency to document
architectural scale – despite the nominal accuracies of these
instruments usually being lower than traditional static LiDAR
solutions – with reduced acquisition time.
Considering the acquisition speed as a crucial parameter for
evaluating the Flash system, the presented research includes a
comparison between a point cloud acquired with a SLAM-based
system and the point cloud derived from the short static scans.
For this reason, the vaulted porch of the cloister has been
surveyed using a SLAM-based mobile scanning system (Stonex
® X120GO) (Martino et al. 2023; Tanduo et al. 2023). A
coloured point cloud – the system is equipped with three 5 MP
cameras in order to provide radiometry to the acquired scans –
has been therefore collected. Subsequently, the data derived from
the employed scanning systems have been co-registered in the
same reference system using (using an ICP-based algorithm) and
compared. The main differences between the used systems – and
the derived primary data – can be identified as follows:
•
Density of the collected point clouds. The point cloud derived
from the MMS is composed by ca. 50 mln of points, while
the Flash point cloud (before proceeding with decimation and
filtering procedures) is composed by ca. 1150 mln of points
(for this comparison, only the scans acquired under the
porch’s vault and covering the same surfaces of the path
followed by the MMS have been considered). The Flash
point cloud is significantly greater and denser than the MMS
one, if the architectural scale application requires higher
detail. However, is often necessary to properly plan a tailored
data acquisition strategy to optimize amount of data collected
and filter and eliminate redundant points, thereby improving
data manageability;
Acquisition time. The acquisition of the point cloud collected
with the Stonex®120GO required approximately 5-10
minutes while, regarding the Flash technology, during the
scanning operations 47 scans were collected and the time
required was approximately 45-60 minutes. Despite the
strategy related to the use of Flash scans being extremely
competitive in terms of acquisition speed compared to a
traditional terrestrial laser scanning method, in this case, the
time required by the Stonex ® 120GO system is significantly
lower. This emphasises how one of the main features that has
made scanning systems belonging to this family increasingly
popular in the field of architectural metric surveying is the
decisive optimisation required time for the acquisition phase.
3D metric accuracy. After a discrepancy analysis performed
between the considered data, it can be observed that 97.6%
of the analysed points are characterised by deviation lower
than ±0.02, evidencing how both point clouds are – from a
metric accuracy perspective – consistent with the
requirements needed for architectural-scale documentation.
(Figure 14).
Level of detail/geometric definition. From (Figure 15) it is
possible to observe how the Stonex system was able to detect
the main surfaces and architectural elements of the porch, but
the level of spatial resolution which characterises the Flash
point clouds is significantly higher – comparable to the one
achieved by traditional static scans – and in comparison, with
the MMS, this data effectively described the details and
elements belonging to the decorative apparatus.
Radiometry. Despite the recent trend in the development of
SLAM-based MMS, which has increasingly moved towards
the possibility of providing radiometry to the point clouds
collected with Flash technology by implementing digital
cameras in the acquisition systems, in this case the visual
comparison between the two-point clouds (Figure 16)
reveals that the Flash point cloud, enriched with images
acquired with the Panocam, provides a significantly better
result in terms of radiometric quality (also visible in Figure4).
Figure 14. Cloud-to-cloud discrepancy analysis between MMS
data and Flash data.
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLVIII-2-W4-2024-381-2024 | © Author(s) 2024. CC BY 4.0 License.
387
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVIII-2/W4-2024
10th Intl. Workshop 3D-ARCH “3D Virtual Reconstruction and Visualization of Complex Architectures”, 21–23 February 2024, Siena, Italy
FARO,
2023.
FARO
“Hybrid
Reality
Capture”.
https://www.faro.com/it-IT/News-Library/2023/FARO-ReleaseHybrid-Reality-Capture.
Jiao, J., Zheng, W.-S., Wu, A., Zhu, X., & Gong, S., 2018. Deep
Low-Resolution Person Re-Identification. Proceedings of the
AAAI Conference on Artificial Intelligence, 32(1).
Figure 15. Visual comparison between the two analysed scans.
(a) Detail of the point cloud acquired using the slam-based
scanner Stonex ® X120GO (portion of the vaulted porch and
elements of the decorative apparatus); (a) Flash scan.
Kim, J. H., Lee, J. S., 2018. Deep residual network with enhanced
upscaling module for super-resolution. IEEE/CVF Conference
on Computer Vision and Pattern Recognition Workshops
(CVPRW), 800-808.
Li, R., Li, X., Fu, C. W., Cohen-Or, D., & Heng, P. A., 2019. Pugan: a point cloud upsampling adversarial network. Proceedings
of the IEEE/CVF international conference on computer vision,
7203-7212
Martino, A., Breggion, E., Balletti, C., Guerra, F., Renghini, G.,
Centanni, P., 2023. Digitization approaches for urban cultural
heritage: last generation MMS within Venice outdoor scenarios. The
International Archives of the Photogrammetry, Remote Sensing and
Spatial Information Sciences, XLVIII-1/W1-2023, 265-272.
Figure 16. Point cloud acquired using (a) the slam-based scanner
Stonex ® X120GO; (b) the Faro Flash.
5. CONCLUSIONS AND PERSPECTIVES
In conclusion, the research presents the exploration of the new
Hybrid Reality Capture™ (HRC), with Flash Technology
(FARO Tech.) as a benchmarking of the scanning system
approaches in indoor/outdoor heritage contexts. It also propose a
comparison with consolidated static FARO scans typology and a
SLAM-based mobile mapping data. The research evaluated and
summarize the main performances of the system in terms of final
3D data, examining in particular the results of the new
upsampling algorithm based on hybrid LiDAR and Panocam data
equipping the scanner. Different factors are related to the
upsampling performance, influencing quality, density and
continuity of the final 3D data: scan position and distance from
the object; LiDAR rays’ incidence on the surfaces and quality of
radiometric data. This is undoubtedly a up-to-date promising
technological improvement in direction of hybridization of
sensors, automation in procedures and speediness of site survey.
Nevertheless, as introduced, the algorithm at the bases of this
such powerful upsampling is still under FARO patent.
REFERENCES
Ajmar, A., Boccardo, P., Broglia, M., Kucera, J., Giulio Tonolo, F.,
Wania, A., 2017. Response to Flood Events: The Role of Satellitebased Emergency Mapping and the Experience of the Copernicus
Emergency Management Service. Flood Damage Survey and
Assessment: New Insights from Research and Practice, AGU,
American Geophysical Union, 229-244.
Bonfanti, C., Patrucco, G., Perri, S., Sammartano, G., Spanò, A.,
2021. A new indoor LiDAR-based MMS challenging complex
architectural environments. The International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Sciences,
XLVI-M-1-2021, 79-86.
Charles, R. Q., Su, H., Kaichun, M., Guibas, L. J., 2017.
PointNet: Deep learning on point sets for 3D classification and
segmentation. Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 652-660.
Nguyen, N. H. T., Perry, S., Bone, D., Le Thanh, H., Xu, M., &
Nguyen, T. T., 2022. Combination of Images and Point Clouds
in a Generative Adversarial Network for Upsampling Crack Point
Clouds. IEEE Access, 10, 67198-67209.
Park, J., Kim, H., Tai, Y. W., Brown, M. S., Kweon, I., 2011.
High quality depth map upsampling for 3D-TOF cameras.
International Conference on Computer Vision, IEEE, 1623-1630.
Riihimäki, H., Luoto, M., Heiskanen, J., 2019. Estimating
fractional cover of tundra vegetation at multiple scales using
unmanned aerial systems and optical satellite data. Remote
Sensing of Environment, 224, 119-132.
Schulter S., Leistner C., Bischof H., 2015. Fast and accurate image
upscaling with super-resolution forests, IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 3791-3799.
Tanduo, B., Teppati Losè, L., Chiabrando, F., 2023.
Documentation of complex environments in cultural heritage
sites. A SLAM-based survey in the Castello del Valentino
basement. The International Archives of the Photogrammetry,
Remote Sensing and Spatial Information Sciences, XLVIII1/W1-2023, 489-496.
Yu, L., Li, X., Fu, C., Cohen-Or, D., Heng, P., 2018. PU-Net:
Point Cloud Upsampling Network. IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), 2790-2799.
Zhang, G., Yang, J., Zheng, Y., Wang, Y., Wu, Y., Chen, S.,
2021. Hybrid-attention guided network with multiple resolution
features for person re-identification. Information Sciences, 578,
525-538.
Zhang, Y., Zhao, W., Sun, B., Zhang, Y., & Wen, W., 2022. Point
Cloud Upsampling Algorithm: A Systematic Review.
Algorithms, 15(4), 124.
Zheng, L., Yang, Y., Hauptmann, A. G., 2016. Person reidentification: Past, present and future. arXiv preprint
arXiv:1610.02984.
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLVIII-2-W4-2024-381-2024 | © Author(s) 2024. CC BY 4.0 License.
388