Photogrammetric software as an
alternative to 3D laser scanning in an
amateur environment
MARKUS WARNE
EXAMENSARBETE
VID CSC, KTH
Fotogrammetrisk programvara som alternativ
till laser 3D-‐skanning i amatörmiljö
Photogrammetric software as an alternative to
3D laser scanning in an amateur environment
Markus Warne
marwar@kth.se
Examensarbete i medieteknik
Handledare på CSC var Vasiliki Tsaknaki
Handledare på CUT var Krzysztof Skabek
Examinator: Haibo Li
Photogrammetric software
as an alternative to 3D laser
scanning in an amateur
environment
Abstract
Photogrammetric software today is at a level where it is accessible to the mainstream public and without larger effort is able to reconstruct digital 3D models from
photographic input. This thesis investigates the performance of photogrammetricly reconstructed models and evaluates them by comparing the results to their
corresponding reconstructed models from a 3D laser scanner with a focus on smaller objects in an amateur environment. The evaluation is performed on four different objects, which are all individually compared to their scanned counterpart. They are compared both with a subjective judgment of quality and by numerically measuring the point-‐to-‐point distance on the models. From the results conclusions are drawn that the methods can produce similar results albeit there are many performance factors
discovered for a good reconstructions with photogrammetry. The properties of the physical object and the quality of the visual input data stand out as the most important factors.
Fotogrammetrisk
programvara som alternativ
till laser 3D-‐skanning i
amatörmiljö
Abstrakt
Den fotogrammetriska programvaran som existerar idag är tillgänglig för allmänheten men framförallt kapabel att återskapa digitala 3D-‐modeller utan större ansträngning. Denna rapport utforskar och utvärderar möjligheterna att återskapa dessa objekt för att sedan jämföra hur dessa står sig gentemot motsvarande återskapade objekt med en 3D laser skanner. Fokus ligger på att se hur mindre objekt kan återskapas i en amatörmiljö. Testerna genomförs på fyra olika objekt genom att först återskapa dessa digitalt m.h.a. fotogrammetri för att sedan jämföra dessa inviduellt med motsvarande modeller återskapade m.h.a. 3D-‐skanning. Utvärderingen sker subjektivt med en bedömning av kvalité men även genom att mäta avstånden från punk till punkt på modellerna. Från resultaten kan slutsatserna dras att det går att nå likvärdiga resultat med fotogrammetri som 3D-‐skanning men dessa beror på ett antal kritiska faktorer. Objektets fysiska
egenskaper samt kvalitén av den visuella data som används framstår som nyckelfaktorer för att lyckas med en bra digitalt återskapad modell.
1 Introduction ... 1
1.1 Goal of the thesis ... 2
1.2 Research questions ... 2 1.3 Limitations ... 3 2 Background ... 4 2.1 Related research ... 4 2.2 Triangulation ... 5 2.3 3D laser scanning ... 5
2.3.1 How laser triangulation sensors work ... 5
2.4 Photogrammetry ... 5
2.4.1 The basics of Photogrammetry ... 6
3 Method ... 8
3.1 Literature study ... 8
3.2 Quantitative evaluation ... 8
3.3 Qualitative observations ... 8
4 Evaluation setup ... 9
4.1 Data generation -‐ Photogrammetric approach ... 9
4.1.1 Photo environment setup and camera parameters ... 9
4.1.2 Photogrammetric processing: Agisoft’s Photoscan ... 10
4.1.2.1 Camera alignment ... 10
4.1.2.2 Dense point cloud ... 10
4.1.2.3 Mesh construction ... 11
4.1.3 Photogrammetric processing: Autodesk’s 123D Catch ... 11
4.2 Data generation -‐ laser scanning approach ... 12
4.2.1 Environment and setup ... 12
4.2.2 Konica Minolta Vivid 9i ... 12
4.2.3 Model reconstruction from point cloud ... 12
4.3 Data comparison ... 13
4.3.1 Alignment ... 13
4.3.2 Measurement ... 13
5 Results ... 15
5.1 Case 1 – Quadric object ... 15
5.1.2 3D print reconstructed with 123D Catch ... 17
5.2 Case 2 – Angel figure ... 18
5.2.1 3D model reconstructed with PhotoScan ... 19
5.2.2 3D model reconstructed with 123D Catch ... 20
5.3 Case 3 – Monkey figure ... 21
5.3.1 3D model reconstructed with PhotoScan ... 22
5.3.2 3D model reconstructed with 123D Catch ... 23
5.4 Case 4 – Wooden cat ... 24
5.4.1 3D model reconstructed with PhotoScan ... 24
5.4.2 3D model reconstructed with 123D Catch ... 26
6 Analysis and discussion ... 27
6.1 Comparing the results of the photogrammetric reconstructions and the laser scanned reconstructions ... 27
6.2 Strengths and weaknesses of the photogrammetric reconstruction software applications ... 28
6.3 The photogrammetric reconstruction process as a whole ... 29
6.4 Can photogrammetry yield similar results to 3D laser scanning when used in an amateur home setting for smaller objects? ... 30
7 Conclusion ... 31
7.1 Future research ... 32
8 References ... 33
9 Appendix ... 35
9.1 Case 1 – deviation results ... 35
9.1.1 Photoscan model compared to 3D scanned model ... 35
9.1.2 123D Catch model compared to 3D scanned model ... 35
9.2 Case 2 – deviation results ... 36
9.2.1 Photoscan model compared to 3D scanned model ... 36
9.2.2 123D Catch model compared to 3D scanned model ... 36
9.3 Case 3 – deviation results ... 37
9.3.1 Photoscan model compared to 3D scanned model ... 37
9.3.2 123D Catch model compared to 3D scanned model ... 37
9.4 Case 4 – deviation results ... 38
1 Introduction
Recently, there has been a growing interest in 3D printing technologies, with a number of hardware and applications with a focus on the context of everyday use (Crum 2014). This development indicates a future where 3D printing could be a tool for personal use for everyone, not just experts. The 3D printing techniques are quickly advancing to become more accessible to the general public with new models created specifically for home use and a lower budget (Matter and Form 2014). What has not been discussed in depth is the other side of the spectrum. What will we print? The models and data must come from somewhere. As the demand and ability to print 3D models increases, the supply of models must also follow according to basic economic theory. When this happens, the market will desire a method for gathering 3D model data that is accessible on an amateur scale to as many individuals as possible, at a low cost.
Methods such as laser scanning are not new, the first triangulation laser scanning
technology was developed already in 1978 (Mayer 1999). However, these methods were and are still inaccessible to the mainstream public, at least to some degree, as the cost of acquiring the technology or the knowledge needed to operate it is simply too high for the average user. A relatively cheap alternative to laser scanning is stereo
photogrammetry. With the help of specific software and the advanced triangulation algorithms that are available today, this method can be used, similarly to a 3D scanner, to reconstruct and digitally model real life physical objects from just a set of ordinary photos. These methods now open up new possibilities when it comes to creating and sharing 3D models on a much larger scale, by amateur users, as they become more accessible from an economic and technological viewpoint to the general public.
1.1 Goal of the thesis
The goal of this thesis is to take a closer look at 3D reconstruction of objects, by using photogrammetry software as an alternative to laser scanning. Furthermore, to
investigate if this is a viable alternative for mainstream users and available to generate models of a comparable quality to that of a reconstructed model from a laser scanner. The main focus will be on comparing the models reconstructed from the two methods mentioned above (reconstruction with laser scanning and reconstruction with stereo photogrammetry), as described in Figure 1. Specifically, I will investigate the deviations of the surfaces between the two methods while trying to assess the quality of the
models, by comparing them to each other numerically.
Figure 1 – Overview of reconstruction methods
1.2 Research questions
• Can photogrammetry yield similar results to 3D laser scanning when used in an amateur home setting for small objects?
o Numerical comparison of deviation between the reconstructed surfaces
• How do these two methods (laser scanning and stereo photogrammetry) reconstructing models compare to each other?
1.3 Limitations
The comparison is limited to two photogrammetric methods and the laser scanning is used just as reference measurement. The quality of the scanned models and accuaracy for the specific scanner has already been carried out in another publication (Spytkowska 2008). Ideally a set of models from several laser scanners would be used as to be able to generalize the results from laser scanning and identify similarities. In addition, only two photogrammetric software applications will be used to reconstruct 3D models, which could also limits the conclusions drawn from comparing the two techniques and identifying general strengths and weaknesses.
Furthermore, the quality of the reconstructed objects will be assessed subjectively, due to the nature of such qualitative evaluations. The limitation is the fact that there is no digital original object to compare and quantify the deviation from the digital
reconstructions, as in most cases the originals are small physical objects.
The reconstructed models will be compared to a digital representation of the original, in this case, a 3D scanned version of the original object, which in turn will be treated as reference for deviation measurements. This will provide quantifiable results of deviation between the two methods, 3D scanning and photogrammetry, but there is no way to compare these to that of the original object.
One of the big limitations of this thesis is that in order to stay true to the “amateur mainstream user” perspective a number of commercial software applications have been used in parts of the process. This generates some “black box” parts where the
transparency of the process is limited, as we only know what we put in and what comes out. Assumptions are made in these cases based upon general procedures and
algorithms within the field but complete certainty can’t be achieved. It has been made clear which these parts are and when assumptions are made they are clearly indicated.
2 Background
2.1 Related research
Laser scanning has historically been used for scanning small objects in a controlled environment, where there is a possibility of scanning the object in 360 degrees angle. This is due to the fact that 3D scanners often have a small field of view and larger objects often have to be scanned in several iterations and then combined together in order to reproduce and complete the model (Chen et al. 2000; Seitz & Curless 2006). However there are some exceptions with laser scanners such as LIDARs which are laser scanners optimized for capturing large geographic areas and similar scenarios. On the other hand, photogrammetry has mostly been used in these types of situations, when there is a need to scan large-‐scale areas, such as air photography, cartography, mapping of
archaeological sites and other situations with large objects, when there is no need to capture small details but rather focus on extracting measurements etc. Photogrammetry is rapidly becoming more and more common, for reconstructing high detailed 3D
models (Blizard 2014; Poznanski 2014). This can be seen especially on the Internet, where its accessibility to the mainstream public has inspired many hobbyists and so called DIY (do it yourself) enthusiasts (Blizard 2014) but also in the entertainment industry such as games (Poznanski 2014) and movies (Wolff 2004). Previously
photogrammetry has also been used in combination with laser scans as a compliment to provide accurate textures for the model that were generated by laser scanning.
Attempts to reconstruct digital 3D models with the help of photogrammetry were already done in 1984 (Benard 1984) and there are even some cases where the results of the reconstructions have been compared to laser scanning techniques (Baltsavias 1999; Fassi & Fregonese 2013). However most of them are, as mentioned earlier, focused on archaeological, architectural or geo-‐data scenarios where the objects are usually very large in scale. This thesis focuses on the perspective of smaller sized objects, up to one cubic meter, and as some earlier research suggests (Baltsavias 1999) this area might be more challenging for photogrammetric reconstruction of models. In addition, 3D
has not been applied in these scenarios as often and if so only partly or complimentary for smaller objects.
2.2 Triangulation
Triangulation is a central concept for both of the techniques applied in this thesis, as both photogrammetry and most laser scanning methods are based around this principle. It is a method used to calculate the position of points in 3D space. It is used in a wide range of applications and scenarios such as navigation, astronomy and many more due to its broad and dynamic origin. Triangulation works by mathematically intersecting converging lines in space from at least two known points to that of the investigated point. By measuring the difference in the angle to the investigated point the precise location of the point in 3D space can be determined. (Spytkowska 2008)
2.3 3D laser scanning
2.3.1 How laser triangulation sensors work
A triangulating laser scanner simplified consists of two components, a transmitter and a receiver. The transmitter usually consists of a laser diode that projects a ray of light on the object. A charged coupled device (CCD) sensor detects the reflection and due to displacements in the object the angle of reflection varies depending on form and distance. Thus the difference can be measured due to the principles of triangulation. From this data, a point cloud is generated based on each specific measurement, each point with a specific distance from the scanner. This measurement is carried out
thousands of times to generate a large point cloud representing the object. This discrete point data cloud can then be interpolated, usually with the help of some complimentary software, to create a 3D surface that consists of not just points but polygons.
(Spytkowska 2008; Dold & Brenner 2006)
2.4 Photogrammetry
Photogrammetry is often applied in topography scenarios like satellite and aerial photography but also in close range scenarios, such as 3D reconstruction.
Figure 2 – Input and output of optical data capturing methods
2.4.1 The basics of Photogrammetry
The core principle for photogrammetry is triangulation. Due to several photographs (at least two or more) with overlapping information, rays, as they are called, can be
photogrammetry applies this principle to multiple points at a time with theoretically no limit to the number of points measured at the same time (Slama et al. 1980).
To be able to compare the 3D laser scanning method to the photogrammetric approach we have to go from physical 3D domain to digital 3D domain as seen in Figure 2.
However strictly photogrammetry consists of 2D input in the form of photos, thus we have to include the photography part from Figure 2 to achieve the complete flow of physical 3D domain to digital 3D domain.
As mentioned before (Figure 2) photogrammetry is in a way the reversed process of photography. Unfortunately the photographic process is not perfect as information is lost when taking a photo, if it was perfect, as in no information lost, just two photos would be more than enough to recreate the 3D scene. So to compensate for this missing information several photographs (absolute minimum of two) have to be used to aid the calculation. The coordinates acquired from these calculations are the final result from photogrammetry, this can be presented in the form of a point cloud or some other data set that then usually is used further to extrapolate a 3D surface (Greve 1996; Schenk 2005).
For best results the images used as input for photogrammetry a few parameters are important. First is the focus of the camera, since photogrammetry uses pixel for points in triangulation blurry images are very tricky for pinpointing the position of features and other elements that are out of focus. Furthermore the resolution is a very important parameter much due to the same reason mentioned earlier meaning more pixels equal more accurate calculations. Consistent lightning is also a factor that helps
photogrammetry in identifying various elements in the photograph. Varying light casts different shadows from picture to picture this can have a destructive effect, as the same feature can have varying intensity and color depending on the image analyzed at the moment (Greve 1996; Schenk 2005).
3 Method
3.1 Literature study
To see what has been done in the field of 3D reconstruction with the use of
photogrammetry a literature study has been done. Google Scholar and KTH Royal Institute of Technology’s “Primo” service have been used as primary knowledge wells for academic research within the area. Furthermore high detail 3D reconstruction from photogrammetry is a relatively new concept relevant information has been found on blogs and other technical news themed websites. Google search and Wikipedia have been used as a compliment or in conjunction to the named sources above.
3.2 Quantitative evaluation
Quantitative methods are applied in this report to measure the difference in distance between points on several reconstructed 3D surfaces. This is done with the help of software that generates points on the reference surface and then tries to match them with the test surface. Quantitative distance data is received for each of these points. Unfortunately only the disparity between two reconstructed objects can be compared quantitatively and not between the reconstructed objects and the original. This is due to the fact that the original is not in the digital domain and any “conversion” from physical to digital is simply another form of reconstruction method. This leaves the “quality” parameter, from original to reconstruction, unquantifiable and a subjective matter. What can be compared are the disparities of the surfaces from the reconstructed models, which is also the main goal in this thesis, to find out if the methods are comparable and yield similar results in a non professional environment.
3.3 Qualitative observations
give relatively small quantifiable differences but the discrepancy in perceived quality between the original object and the reconstructed model may be very large.
4 Evaluation setup
The photogrammetric evaluation process consists of two major parts, the first is the reconstruction of the physical object in 3D domain – this will be called data generation. The second part consists of analyzing the reconstructed objects and comparing them to each other. This is where the quantified data is extracted from the objects – this will be called data comparison.
As the software applications used in this thesis are commercial, the specific algorithms and techniques used are not available to the public domain. Therefore a general
approach to reconstruction with photogrammetry and laser scanning has been
described and for the specific software that follows all information that is available for each application will be presented and assumptions based on what is known.
4.1 Data generation -‐ Photogrammetric approach
4.1.1 Photo environment setup and camera parameters
The data generation environment was set up in a home setting with varying lightning conditions as to stay in line with the thesis main goal of analyzing the possibility for mainstream public to apply this technique in a non professional context. A tripod was used to aid the stability and positioning of the camera and the lightning consisted of mixed indoor light. The camera was placed as close as possible to the object for the majority of the picture to cover most of the image area to preserve as much detail as possible.
As mentioned before resolution and focus are two very important parameters for the reconstruction process and these have been taken into consideration in this setup. However consistent lightning is also an important factor for best results from
photogrammetric reconstruction but there has been no effort in this setup to minimize this. This is to mimic home conditions of mainstream users to stay in line with the thesis goal and as in a home environment there is usually no way to achieve evenly lit objects such as with studio lightning conditions. However it is assumed that a good depth of field and focus can be achieved in a home environment.
4.1.2 Photogrammetric processing: Agisoft’s Photoscan
The procedure of photograph processing and 3D model construction is described with the following four stages according to the PhotoScan manual (Agisoft LLC 2011).
4.1.2.1 Camera alignment
The software searches for common points in the collection of photographs. To be able to match them, the software also calculates the position and orientation of the camera for each picture. It is very probable that this is carried out with some form of triangulation described earlier. The result of this process is a sparse point cloud where several points have been identified and matched over the different camera positions. It is important to have in mind that several points can be calculated and matched per camera location (a single photo). However this sparse point cloud is not what the reconstruction of the 3D model is based on unless explicitly specified by the user. The information that is used further in the process of reconstruction is mainly the set of camera positions gathered in this stage, presumably with the intention of using them as a starting point for further triangulation of points.
4.1.2.2 Dense point cloud
4.1.2.3 Mesh construction
The third stage consists of constructing a 3D polygonal mesh representing the object surface based on the dense point cloud.
In most cases two algorithmic methods are available that the software can apply for 3D mesh generation:
• Height field – Optimized for modeling of planar surfaces, such as terrains or bas-‐ reliefs. For aerial photography processing, it requires a lower amount of memory and allows for larger data sets processing.
• Arbitrary -‐ For closed objects, such as statues, buildings, etc. It doesn't make any assumptions on the type of the object modeled, which comes at a cost of higher memory consumption.
In this thesis the arbitrary algorithm method was used for all cases of photogrammetric reconstruction processing.
Once the mesh is constructed, the user has the ability to somewhat edit it. Non-‐complex corrections such as mesh decimation (simplification), removal of detached components and the closing of holes can automatically be performed by the software in the mesh generation process.
Furthermore the final stage of the model construction is the application of textures, as this is not relevant in this thesis it will be ignored.
4.1.3 Photogrammetric processing: Autodesk’s 123D Catch
Unfortunately very little information can be found about 123D Catch’s reconstruction process. Presumably the process is somewhat similar to that of PhotoScan. It can also be assumed that again triangulation is a key algorithm and that it is based on characteristic points identified in the set of pictures.
The practical process is very similar to PhotoScan’s. The user takes a set of overlapping photos of the object from different angles. The difference is the processing is done via the software applications server. The user uploads the photoset to the server and after a while, depending on image size and quantity the server returns a reconstructed 3D model.
This is both an advantage and a flaw comparing it to PhotoScan. By automating the process only the photographs have to be provided, this can be helpful in the aspect of a mainstream user and it can be assumed that the server has better computational power than if the reconstruction would be performed locally. However this limits the amount of influence the user has on the reconstruction such as key parameters for reconstruction and insight into the process and therefore also the end result.
4.2 Data generation -‐ laser scanning approach
4.2.1 Environment and setup
The scans were carried out in Gliwice, Poland at the Institute of Theoretical and Applied Informatics (part of the Polish Academy of Sciences) in one of their offices. The laser scanner was a Konica Minolta Vivid 9i and it was connected to a mechanically rotatable platform, which together with the scanner were operated through the complimentary native software for the scanner.
For the scanner to encompass the whole object several scans had to be carried out from different angles, this was controlled with the rotatable platform. The angle of rotation per scan was 30° up to a complete circle of 360° resulting in 12 overlapping scans.
4.2.2 Konica Minolta Vivid 9i
The scanner used for the conducted tests is a Konica Minolta Vivid 9i. The Vivid 9i is a scanner created for small to medium sized objects which it can capture with high detail. This is a triangulating laser scanner with a charged coupled device (CCD) receiver like the ones described in chapter 2.3.1.
4.2.3 Model reconstruction from point cloud
pieces have to be aggregated into a single surface to complete the process. This was done by aligning each individual piece in respect to the others and merging them into a single shape to construct a 3D model. The software application used for this was
Geomagic Qualify which is used to measure, align and compare fabricated parts to their digital blueprint in the production industry.
4.3 Data comparison
4.3.1 Alignment
Alignment is a very important part of the measurement process, as the objects need to be identically aligned for the distance results to be accurate. This is in turn a problem as the objects vary in shape and form, it becomes a subjective judgement if the objects are aligned properly. First n-‐point alignment is used for a rough alignment of the objects. This is based on a physical person identifying common points on both objects and marking these. The minimum for this alignment is 3 points but can be as many as one would like, therefore the name n-‐point alignment. When the points are identified the objects are matched up so the points on both objects align with eachother.
After this rough alignment a global registration algorithm is used. Randomly selected points on the reference surface are used to reposition the test object to minimize the overall distance measured from the points. This somewhat eliminates the subjective factor from the alignment process but the first alignment step is still based on human perception.
4.3.2 Measurement
Due to the surfaces being different in form and quality it is not always possible to map a point on the reference surface to a corresponding location on the test surface. In these cases the difference can’t be measured between the two points. However it is fair to assume that the deviation in the current point in this case is larger or similar to the maximum deviation of the measured points.
5 Results
In this section there are four cases presented with varying objects in complexity and size to represent a wide spectrum of smaller objects that may be found in a home
environment for amateur 3D reconstruction. For each case a view of the original and all reconstructions is shown with a description of each surface. Furthermore color maps are presented from four different perspectives of the deviation between the compared surfaces. The deviation is always measured in millimeters and the legend of the color map shows that green areas are the ones closest to zero in deviation. Further towards red are larger positive deviations and towards blue larger negative deviations.
5.1 Case 1 – Quadric object
This first case is a type of calibration case different to the other test cases that follow in this section. The original object here is a digital 3D model of a simple quadric surface (A). The digital original was then reproduced into the physical domain (scale 1:1) with the help of a 3D printer (B) to then again get reconstructed to digital domain (C&D) with the help of the photogrammetric methods presented earlier. The physical object is small and measures 100x95x45 mm (height x width x depth).
The interesting thing here is that we have a digital original giving us some way of comparing the deviation between the original and the reconstructed models. This is not possible in the other cases, as the original object is physical.
Figure 3 -‐ A) Digital original 3D model. B) Physical 3D-‐print based on A. C) Reconstructed digital model with PhotoScan based on B. D) Reconstructed digital model with 123D Catch based on B.
5.1.1 3D print reconstructed with PhotoScan
What is presented here are four different perspectives (top, bottom, isometric and side view), of the deviation between the original (A) and the reconstruction with the help of PhotoScan (C).
Figure 4 – Case 1, 3D scan compared to reconstruction with PhotoScan
The numerical difference between the digital original and the reconstructed model is very small. As shown in the full table of points in appendix chapter 9.1.1, about 96% of the points are placed in the interval of -‐0,1386 to 0,304 mm. So the numerical difference over 96% of the surface is less than half of a millimeter.
Deviation (mm)
Max. Upper
Deviation Max. Lower Deviation Average Deviation Standard Deviation
5.1.2 3D print reconstructed with 123D Catch
The same perspectives as in the previous comparison are also displayed here. This is the comparison of the digital original (A) and the reconstructed model with the help of 123D Catch (D).
Figure 5 – Case 1, 3D scan compared to reconstruction with 123D Catch
The results are similar to the reconstruction with PhotoScan (C) and again the numerical difference here is small, although a bit larger than the results from PhotoScan. What we can see here is also that the differences gravitate a bit more towards a negative distance difference whereas when reconstructing with PhotoScan we see the differences
gravitating towards a positive difference.
5.2 Case 2 – Angel figure
In this case the original object is a physical figure of an angel, which consists of a mix of small soft and hard features with an overall complex surface. The object is small and measures 80x50x40 mm (height x width x depth). The comparison is conducted between the photogrammetricly reconstructed models (C&D) from the two different methods and compared to that of a reconstructed digital model from a 3D scanner (B).
Figure 6 -‐ A) Physical original. B) Digital 3D-‐model reconstructed with the Vivid 9i based on A. C) Reconstructed digital model with PhotoScan based on A. D) Reconstructed digital model with 123D Catch based on A
5.2.1 3D model reconstructed with PhotoScan
Here we see four different perspectives (front, left, back and right), of the deviation between the 3D-‐scanned reconstruction (B) and the reconstruction with the help of PhotoScan (C).
The color map does not classify the grey areas in this comparison. This is due to a too large discrepancy between the two compared surfaces in the specific area and the deviation measurement algorithm cannot identify the corresponding point on the other surface. Simply put a point on surface C does not match any point on surface B and therefore the difference cannot be measured.
Figure 7 -‐ Case 2, 3D scan compared to reconstructed model with PhotoScan
The reconstructed model is very noisy and distorted, even so we see that the numerical difference over most areas of the surface is still small, only 0.5 mm. With that said it is important to acknowledge the grey areas where points were unable to match we can assume the difference is greater than +/-‐5 mm.
5.2.2 3D model reconstructed with 123D Catch
The same perspectives (front, left, back and right), of the deviation between the 3D-‐ scanned reconstruction (B) and the reconstruction with the help of 123D Catch (D). As previously the color map does not classify the grey areas due to the same reasons in 5.2.1.
Figure 8 -‐ Case 2, 3D scan compared to reconstructed model with 123D Catch
Although both photogrammetric reconstructions contain artifacts and some noise the key thing to notice here is that there are larger areas on the test surface (D) which can’t be matched to the reference surface (B) than in the previous comparison, C vs B. This has a significant impact on the key values like average deviation as pieces of the dataset are missing.
5.3 Case 3 – Monkey figure
As in the previous case the original object here is a physical figure of three monkeys, which consist of mostly small sharp features. The overall surface of the object is very complex with a high amount of detail. The object measures 70x105x50 mm (height x width x depth). A comparison is performed between the photogrammetricly
reconstructed models from the two different methods (C&D) and compared to that of a reconstructed digital model from a 3D scanner (B).
5.3.1 3D model reconstructed with PhotoScan
Four perspectives (front, left, back and right), of the deviation between the 3D-‐scanned reconstruction (B) and the reconstruction with the help of PhotoScan (C) are shown here.
Figure 10 – Case 3, 3D scan compared to reconstructed model with PhotoScan
This statue has a similar size comparable to Case 2 yet the discrepancies here are significantly smaller. The deviation is also very evenly spread over the whole object suggesting there is little to no noise.
5.3.2 3D model reconstructed with 123D Catch
The same perspectives (front, left, back and right). The deviation here is larger than in the comparison of B vs C and the deviation is not as evenly spread out as in 5.3.1.
Figure 11 – Case 3, 3D scan compared to reconstructed model with 123D Catch
Deviation (mm)
Max. Upper
Deviation Max. Lower Deviation Average Deviation Standard Deviation
5.4 Case 4 – Wooden cat
This case is unique in the sense that the object here is significantly larger than the previous cases. The object measures 410x150x50 mm (height x width x depth). As before in the previous cases the original object here is a physical figure of cat carved in wood that consists mostly of large soft features and is quite simplistic in shape. The comparison is conducted between the photogrammetricly reconstructed models from the two different methods (C&D) and compared to that of a reconstructed digital model from a 3D scanner (B).
Figure 12 -‐ A) Physical original 3D model. B) Digital 3D-‐model reconstructed with the Vivid 9i based on A. C) Reconstructed digital model with PhotoScan based on A. D) Reconstructed digital model with 123D Catch based on A.
5.4.1 3D model reconstructed with PhotoScan
Figure 13 – Case 4, 3D scan compared to reconstructed model with PhotoScan
5.4.2 3D model reconstructed with 123D Catch
This reconstructed model is not as noisy as the previous case and the discrepancies are much more evenly grouped. Again even though some areas are quite deviant most of the surface has a very small deviation of 0.5 mm.
Figure 14 – Case 4, 3D scan compared to reconstructed model with 123D Catch
6 Analysis and discussion
These comparisons are carried out between the 3D laser scan of the object and the photogrammetric reconstruction so small numerical differences do not always
correspond to an ideal reconstruction as errors might very well exist in the laser scan of the object (Spytkowska 2008). This is not a point that has or will be explored further in this thesis, as the goal is only to compare the methods with each other and not with the physical original on a quantifiable level. As explained earlier this is not possible in most cases and neither the goal of this thesis. Nevertheless it is important to keep that in mind when analyzing these results. We cannot assume the digital laser scan of the object is ideal but it is the reference for comparison of the photogrammetric reconstruction methods.
6.1 Comparing the results of the photogrammetric
reconstructions and the laser scanned reconstructions
Although the numerical differences in all these cases are by comparison relatively small excluding some minor areas it is important to note the difference in perceived quality of the objects. Looking at the results, the conclusion that lower average deviation also correlates to a better-‐perceived quality can be said to be true. Furthermore we can see that the more evenly the deviation is spread out over the surface the better the
perceived quality of the reconstruction. Even if large areas are quite off, the key factor affecting the perceived quality is noise, and large areas with similar deviation often suggest small amounts of noise or at least a uniform distortion. This of course assumes that the 3D laser scanner has reconstructed the physical original in a satisfying manner, which is assumed here as discussed earlier, since it is these two surfaces we are
comparing and not the original. These conclusions were expected but cannot be said to always be true since this is a subjective judgment and the perception of quality might vary for each individual.
6.2 Strengths and weaknesses of the photogrammetric
reconstruction software applications
It is also interesting to try and analyze the strengths and weaknesses of the algorithms of the two photogrammetric software applications in their ability to reconstruct the surface of the object. In general we can see that the algorithms struggle when faced with areas and whole surfaces that lack large or multiple changes, in other words smooth soft features and same colored surfaces as seen in case 2 and 4. The type of change whether it is surface based or color based seems to be of less importance. As we can see in case 1, the simple and smooth surface of the quadric is compensated by the complex pattern painted on the surface. Furthermore surface differences like the fur of the monkeys in case 3 also result in the same type of change such as color changes due to lightning giving the features a shadow and thus a different intensity. Therefore changes in either surface or color result in a similar difference on a photograph and it can be assumed that this is what the algorithms use to identify points of measurement for the triangulation as mentioned in chapter 2. This property will be named Δ (Delta); objects with a high amount of Δ produce better results when reconstructed.
The cat, case 4, is a good example where we see quite small Δ, as changes in color and surface over large areas contribute to confusion for the photogrammetric applications as they most likely have no way of identifying the points on the surface and therefore struggle aggregating the orientation of the respective images.
We can also come to the conclusion here that Δ is not the only factor for good
photogrammetric reconstructions but also the density of the change. If we look at case 2, the angel figure has some areas with soft round features (low Δ) and some areas like the hair with high Δ. The areas with high Δ have been reconstructed quite well however the lack of surface nearby with high amounts of Δ create a challenge for the reconstruction algorithm and the result is quite poor. The conclusion that can be drawn from this is that even if an object has a high amount of Δ it has to be spread out over the whole surface for a good reconstruction, concentrated amounts will only give a good result in that specific area. The density of Δ will be called ρ (rho).
It is also clear that there is a difference in the algorithms between the two applications. PhotoScan’s reconstruction generates a substantial amount of noise but still manages a quite small overall numerical deviation over the whole surface, suggesting a more intense frequency of points triangulated or a large difference in the interpolation between points. In the case of 123D’s reconstruction the points calculated are probably fewer, this contributes to, in most cases, that the perceived quality is much higher due to the lower amount of noise and it still captures the overall features. However it’s
important to note that in the best reconstruction, case 3, PhotoScans point frequency generates a more accurate reconstruction by capturing far more detail than 123D Catch thanks to the larger amount of points calculated, although as mentioned above, this is not always an advantage as errors can become plentiful.
Another interesting difference between the algorithms is that PhotoScans deviation is usually positive, differences between the reconstructed surface point outwards from the object however 123Ds deviation is often negative, pointing inwards from the surface. This results in PhotoScans reconstructions generally ending up larger in size than their equivalent laser scan and 123D reconstruction ending up smaller than their counterpart. This is presumably a side effect of each applications calculation algorithm and it is hard to know if it is related in any way to the perceived quality of the object although it can have some impact when trying to recreate objects made to scale.
6.3 The photogrammetric reconstruction process as a whole
Photogrammetry, as can probably derived from just the name, is highly dependent on the input data, in other words the photographs quality. This is not the same subjective qualities of the human perception of a “nice” picture, but rather good quality as in
maximum preservation of unaltered visual information as described in chapter 2.4.1 and 4.4.1. This is the main bottleneck for this technology and as photographs with altered or low amount of visual information also give bad results when processed by the
reconstruction algorithm. Bad input can be somewhat mitigated with the help of high amounts of Δ and ρ on the object but only to some extent. This thesis focus was more on the results of photogrammetric reconstruction and comparing those to that of
