E3sconf Gi2021 04001
E3sconf Gi2021 04001
E3sconf Gi2021 04001
1051/e3sconf/202122704001
GI 2021
Abstract. Four open source and three commercial software been used to
create point clouds from identical data: images taken at five different
locations. The investigation compares elapsed time and the size of the
generated point clouds for the different software. Based on the tests,
suggestions for users are provided.
1 Introduction
Nowadays photogrammetry and laser scanning (LiDAR) competes in generating point
clouds. The present study point clouds generated with photogrammetry is discussed. The
aim of the study is to provide an overview of the available commercial and open source
software, and also deliver a comparative analysis on their efficiency based on certain
aspects. Among the huge number of available software, seven has been selected for the test,
which are considered to be most widely used ones.
*
Corresponding author: lehomate@gmail.com
© The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons
Attribution License 4.0 (http://creativecommons.org/licenses/by/4.0/).
E3S Web of Conferences 227, 04001 (2021) https://doi.org/10.1051/e3sconf/202122704001
GI 2021
coordinates of the reference points (a subset of key points, which can successfully be
determined by the orientation).
The resulted point set is the so-called sparse point cloud, the outcome of the SfM. A
sparse point cloud may visualize appropriately the texture, geometrical content is rough
(e.g. edges). Even though the SfM derives a 3D model, it does not generate new corners or
edges, but applies the edges of the texture [3].
After the SfM, the dense point cloud is generated. Most of the software use telemetry data
or the absolute oriented point cloud to generate derived products, such as orthophoto,
Digital Terrain Model (DTM) or Digital Surface Model (DSM) [4]. The source of the
images can be hand-held camera or cell phone, UAV-borne camera, but also pictures
available on social media (e.g. Flicker, Mapillary) can be used (c.f. Fig 1). Mapillary is a
project dedicated to data acquisition from social media, which resembles to Google Street
View, though Mapillary contains images of front-facing cameras only. The data is collected
volunteers mostly with cell phones mounted to the windscreen of cars. The processing
software may also make use of the EXIF (Exchangeable Image file Format) content of
JPEG files, such as telemetry data, focal length, etc. Generally, 80% or more overlap of the
images is expected.
2
E3S Web of Conferences 227, 04001 (2021) https://doi.org/10.1051/e3sconf/202122704001
GI 2021
Fig. 2. Identified key points, with the accuracy (size of the circle) and the direction of the point on
another image (marked radius of the circle) is displayed. [5]
There are expectations on the Hardware capacities for most of the software, such as
number of processors (CPU) and nodes; often 6-8 nodes may be demanded. The different
steps of the photogrammetric processing sequence mean different workload for the
processors as well. The computation time of some steps depends only on the clock speed,
while others can make use of the multiple nodes. For example, generating a dense cloud is
efficient with multiple nodes, while building a texture or a mesh model needs generally
high clock speed but less nodes. Also, the processing power of the graphics card (GPU)
may also be involved in the processing, for example the computation time of densification
of the point cloud depends notably on the number (and the type) of GPUs. Since this step is
the most time consuming, it is important select all the available GPUs to be used for
processing. Approximately, a graphics card with an additional GPU can reduce the
computational time by 20-25%, which with one more GPU can be further reduced by
~10%. Additional GPUs, however, won't affect the computation time relevantly (1-3%
only), so it does not verify the additional costs.
In case of hundreds or thousands images are processed, the memory (RAM) capacity
should be maximized (at least 64GB is needed). The memory demand depends on the
number of images, and the demanded resolution. As the use of RAM is approximately
linear, by a test run for a subset of data, the computation time can properly be estimated.
Normally, generating the dense point cloud and the mesh model is the most RAM-
consuming. For processing aerial images, in case of some software there are special settings
for the nadir images, which enables more efficient processing. In case of oblique images,
however, the demand on high memory capacity is unavoidable.
Furthermore, data storage is also a relevant issue. As primary drive, SSD is suggested to
be used, which contains both the operational system and the processing software. High
speed SSD enables fast loading and handling of the data.
All in all, for professional photogrammetric processing the optimal computer can be
defined somewhere between the Gamer PCs and laptops. For smaller tasks and study,
middle range PCs can also be sufficient. Finally, the option of online processing should also
be mentioned, as cloud computing is more often supported by photogrammetric software.
3
E3S Web of Conferences 227, 04001 (2021) https://doi.org/10.1051/e3sconf/202122704001
GI 2021
Within the frame of the present analysis, 4 open source (VisualSfM, Regard3D,
COLMAP, ODM/WebODM) and 3 commercial (3DSurvey, Agisoft Metashape, Recap
Photo) software are compared.
3 Used software
For the tests, altogether 7 software has been used [6]. In the following section these are
introduced, and their main parameters are summarized in Table 1.
3.1.1 VisualSfM
One of the earliest open source photogrammetric processing software, which applies
SfM and MVG algorithms. The development has been started in 2011, and no new version
has been published in the last two years. It has a special licence, indicating that it can only
be used for private and research use [7].
For determining the key points, it requires ATI/nVidia/Intel GPU with at least 1 GB
memory. For the 3D reconstruction nVidia CUDA or the CPU can also be used. For
generating the dense point cloud PMVS/CMVS patch is used. No derived products (e.g.
orthophoto) can be generated. Both graphical and command-line interface can be used.
Though the program can efficiently process the images, also the user interface can
easliy be acquired, drawbacks are that the updates have been ceased, and also that
commercial use is not supported.
3.1.2 Regards3D
It has been started to be developed in 2015. Very user friendly, it guides the user with
clear instructions through the processing steps in an interactive manner, enabling setting for
different algorithms and parameterization in dialog boxes. Within one processing sequence,
plenty of variants of processing can be run in parallel. No special requirements for the
hardware is defined, however, it cannot process more than some hundred images at once. It
uses MIT-licence, which enables wide range of use, including its use within a commercial
software as well.
No reference points can be involved into the processing; accordingly, only relative point
cloud can be generated. In order to transform to an absolute coordinate system, some other
software (e.g. CloudCompare) should be used. The software is frequently used in cultural
heritage management for modelling buildings and sculptures [8]. It is a good starting point
for beginners to get acquainted with this technology.
4
E3S Web of Conferences 227, 04001 (2021) https://doi.org/10.1051/e3sconf/202122704001
GI 2021
3.1.3 COLMAP
The COLMAP software has been started as a PhD research, it's first licence has been
published in 2016 as a BSD-license. The main purpose of the software was pure research,
development, and test of algorithms. As a result, the program has a very wide range of
options for parameterization, which may be a difficulty for unexperienced users. It is rather
for research purposes for experimenter users.
Large (more than 10000 images) can also be processed, although the computational
time is not optimal. It can used with graphical and command-line user interfaces as well.
The present version cannot be used for generating orthophoto. Even though the software is
mainly developed by one person, dozens of developers contribute to his work.
5
E3S Web of Conferences 227, 04001 (2021) https://doi.org/10.1051/e3sconf/202122704001
GI 2021
3.2.1 3D Survey
The 3D Survey has been started to be developed back in 2011 by a Sloven group of
(mainly) surveyors. The software has been first published in 2014, since then its popularity
is raising year by year. The company consists of 11 persons, among them 3-5 are purely
working on the development. Its success is based on the user-centric and dynamic
development policy and adaptive support service. Quite a few setting and parameterization
option can be set by the user, which is both an advantage and a disadvantage. Nevertheless,
it is an optimal choice for a beginner user.
There are a number of constantly evolving modules, such as more and more refined
CAD-functions, point cloud classification, volume calculation, automatic profile- and cross
section generation. A unique service of the software is the automatic reference point
recognition, for which special, round-shaped markings are provided by the company. The
recognition, which is based on machine vision, can notably reduce the manual workload.
One of the most commonly used photogrammetric processing software, originated from
Russia started to be developed in 2006, first published in 2010 under the name Agisoft
Photoscan. A huge range of options for parameterization and settings, which also results in
an increased number of source of errors. In case a limited capacity hardware is available, it
can be used by distributing the task into smaller units, so-called chunks: the processing can
be done separately for each chunk, and their outcome can be merged into one. A special
feature is that task lists can be defined for a certain processing sequence, yielding an
automatization of the processing. Different, sometimes quite special camera calibration data
can also be involved in the processing, e.g. shutter compensation. Multispectral images can
also be processed by assigning for each channels a colour. This software has one of the best
point cloud classification method (Fig 4).
6
E3S Web of Conferences 227, 04001 (2021) https://doi.org/10.1051/e3sconf/202122704001
GI 2021
ODM, Lin, Docker ++ A CLI, web Open GPL Pc, Tx, Op, Mm,
WebODM DSM
Agisoft Lin, Win, +++ A, T GUI Commercial Pc, Tx, Mm, Op,
Metashape OSX DSM
7
E3S Web of Conferences 227, 04001 (2021) https://doi.org/10.1051/e3sconf/202122704001
GI 2021
8
E3S Web of Conferences 227, 04001 (2021) https://doi.org/10.1051/e3sconf/202122704001
GI 2021
References
1. J. L. Schönberg, J. Frahm, 2016 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), Las Vegas, NV, 4104-4113. DOI: 10.1109/CVPR.2016.445
(2016)
2. D. G. Lowe, International Journal of Computer Vision, 60, 91–110. DOI:
10.1023/B:VISI.0000029664.99615.94 (2004)
3. S. Bianco, G. Ciocca, D. Marelli, J. Imaging, 4, 98. DOI: 10.3390/jimaging4080098
(2018)
4. F. Remondino, M. G. Spera, E. Nocerino, F. Menna, F. Nex, Photogrammetric record,
29, 146, 144-166. DOI: 10.1111/phor.12063 (2014)
5. M. Lehoczky, Z. Siki , Geodezia es Kartografia 2020/2, 23-27. DOI :
10.30921/GK.72.2020.2.4 (2020)
6. G. Vacca, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-4/W14, 239–
245. DOI: 10.5194/isprs-archives-XLII-4-W14-239-2019 (2019)
7. C. Wu, link: http://ccwu.me/vsfm/ ; last retrived: 20 September, 2020 (2020)
8. H. Rahaman, E. Champion, Heritage, 2, 1835-1851. DOI: 10.3390/heritage2030112
(2019)