The Structure-from-Motion (SfM) technique for 3D modeling has been extensively utilized across numerous application domains, primarily due to its high degree of automation enabled by advanced computer vision methods, which contribute to the development of efficient workflows. Figure 1 illustrates a summarized overview of the standard SfM workflow.

The process begins with the acquisition of a series of images from multiple viewpoints, ensuring significant overlap between consecutive captures. Feature detection is performed on each image, followed by the matching of these features across the entire image set. The SfM phase then employs these redundant feature points to execute Bundle Adjustment, which, through the incorporation of control points or scale bars, generates a sparse point cloud that approximates the geometry of the object or scene ( James et al. 2017). This sparse representation is further refined during the Multi-View Stereo stage, where additional feature detection and matching techniques are applied to produce a dense point cloud. The result is a high-fidelity three-dimensional reconstruction of the captured object or scene ( Luhmann et al. 2020).

As demonstrated in the workflow, the set of input images is integral to the feature detection stages and is repeatedly utilized during the Multi-View Stereo (MVS) process. Consequently, the characteristics and quality of these images, which may vary in configuration and parameters, are crucial in determining the effectiveness and accuracy of the 3D modeling process. According to Hafeez et al. (2018), the comprehensiveness and precision of the resultant 3D models, facilitated by SfM, are intricately tied to both the quantity and quality of discernible points of interest within the images. Therefore, the assurance of high accuracy of the derived products requires the set of images to adequately portray the scene under consideration with an elevated level of detail and quality.

Image quality is inherently related to the lighting attributes of the captured scene, as argued by Dauvin et al. (2018). The authors highlighted the importance of uniform lighting throughout the region of interest in the scene to facilitate high-quality camera calibration, 3D modeling processes, and more detailed photographic capture of the surface of an object or scene. This factor enables the identification of points of correspondence between images and increases the accuracy and stability of the 3D reconstruction. Therefore, complementing scene capture with adequate lighting is fundamental to improving the effectiveness of element detection for objects characterized by intricate surface detail with variations in texture and tone. In many cases, solar illumination alone may be insufficient for objects with sparsely detailed surfaces.

Numerous studies in different domains, including those by Capéran et al. (2012), Kwak et al. (2013), Mishra et al. (2017), Nietiedt et al. (2020), and Nielsen et al. (2023) have addressed the challenge of low object detail using artificial textures projected or applied to object surfaces. This approach has yielded remarkable results, enhanced surface detail, and introduced new patterns, thereby facilitating the detection of a larger number of salient points of interest.

Besides lighting and surface texture, the choice of storage formats is another factor that influences the representation of the imaged object and JPG or JPEG (Joint Photographic Experts Group) format is preferred in widely used capture systems. However, whereas it offers a wide range of colors with minimal storage requirements, its data compression process can result in loss of information or reduced image quality. Such data loss adversely affects the image correlation processes that are essential to the effectiveness of the SfM technique.

To address the quality degradation caused by JPG, some professionals have used photographic equipment that produces uncompressed digital images, known as RAW files. Such files preserve the fidelity of the data captured by the camera sensor, enabling a more accurate representation of the scene. However, the superior quality of RAW images, often stored in TIFF (Tagged Image File Format), results in significantly higher storage space requirements compared to JPEG. Morgan et al. (2017) and Verma and Bourke (2019) investigated those file formats. The former study reported consistency in model quality between them; however, the raw format offered advantages in facilitating extensive image processing due to a richer data representation. Analyses conducted in the latter study revealed a slight superiority in positional accuracy of models generated in the raw format attributed to enhanced detection of salient points of interest. Notably, a more detailed scene representation can potentially yield more effective and higher-quality 3D models.

The photographs were taken in the Geomatics Laboratory of the Transport Engineering Department of the São Carlos School of Engineering at the University of São Paulo. The selected test site has both natural and artificial lighting, which, according to the objectives of this study, influenced the photographic capture settings.

Samples of three materials, namely concrete, metal, and wood, simulated elements commonly used in laboratory testing of structural beams. The dimensions of the objects were 140 cm × 30 cm × 4 cm for the concrete beam, 140 cm × 40 cm × 1 cm for the metal structure, and 140 cm × 18 cm × 4 cm for the wooden beam.

The regions of interest for each object were defined as the surface planes of the respective objects in the captured photographs. Eight sets of acrylic rulers with checkerboard patterns were randomly placed around these regions to facilitate the automatic detection of reference points by the modeling software.

These acrylic plates of known dimensions functioned as scale bars for calibrating and scaling the resulting 3D models. Another set of three acrylic rulers, to be used as control bars (CBs), with checkerboard patterns and known dimensions, were placed within the region of interest of the test specimens to verify the quality of the 3D modeling. The bars were placed at different lengths and positions relative to the axes of the specimens being analyzed. Figure 2 shows the regions of interest and the configuration of SBs and CBs for each material.

Because of its significant dimensions and weight, a robust support structure was necessary to ensure the metal specimen’s stability and safety during imaging. This arrangement resulted in the support structure and the metal specimen being fixed in the center of the room. Consequently, the other specimens were placed next to the metal specimen, creating a consistent environment for analyzing all variables in this research.

In laboratory tests, the standard lighting of the location can cause variations in the color of the same region in different images due to the internal environmental conditions of the facilities. Farhadmanesh et al. (2021) addressed how such variations in representation can negatively affect the effectiveness of the interest point detection (SIFT-like algorithms), as discussed in Badano et al. (2015) and Lurie et al. (2017).

To meet the light quality requirements of this study, two auxiliary lighting units (softboxes) were used, each with a 7,000-lumen LED lamp and a color temperature of 5,000 Kelvin. The softboxes were arranged in three different positioning configurations relative to the specimen to simulate different structural test environments (see Figure 3).

The initial “Standard” configuration ( Figure 3a) depicts the environment in its natural state, utilizing only the room’s ceiling lighting and without any additional light sources. In the “Vertical” model ( Figure 3b), auxiliary lighting was strategically positioned along the camera’s line of sight and directed towards the specimen to illuminate the object vertically and avoid shadowing caused by the acquisition process.

In the “Adjacent” configuration ( Figure 3c), softboxes were positioned in the lateral regions of the test object, simulating tests where other equipment is required between the camera and the object to avoid shadowed areas due to obstruction of lighting.

Finally, the “Beneath” configuration ( Figure 3d) shows an arrangement in which softboxes need to be positioned below the object for replicating experimental environments with space limitations for positioning lighting sources in “Vertical” or “Adjacent” configurations relative to the object.

Reconstruction sets were generated for each lighting scenario on multiple specimens to evaluate the impact of the proposed lighting configurations on the quality of the resulting 3D models. Given that the primary objective of this study is to enhance the image capture process, no post-processing techniques were employed to adjust colors or lighting.

This approach was chosen to ensure proper light exposure of the objects and environment, thereby facilitating the capture process and producing images rich in detectable and correlated features for the SfM and Multi-View Stereo algorithms, as discussed in O’Connor (2018) and Pena-Villasenin et al. (2019) without relying on labor-intensive post-processing steps.

The combinations utilized are presented in Table 1, which highlights the adopted lighting configurations, and the number of images generated at each acquisition stage.

Artificial texture patterns were applied to the surface of the specimens to enhance the detection of points of interest in the materials used, highlighting details of natural texture and color contrast on the object, thus facilitating the detection and correlation of these points between photographic sets.

In Reiss and Tommaselli (2011) and Detchev et al. (2014) image projectors were used to create random patterns on the surfaces of the analyzed objects. However, in the present study, due to the use of reference bars with checkerboard patterns and the analysis of the use of auxiliary lighting, image projection would introduce challenges in the automatic detection of reference bars and would impact the lighting configurations of the environment. Therefore, two random patterns, a checkerboard pattern (T1) and a more complex pattern (T2), were drawn on the surface of the objects. Figure 4 shows the sets of natural textures of the samples and the artificial patterns used.

Due to the rich surface texture of the concrete and wood samples, a 4 cm x 4 cm checkerboard pattern (T1) drawn in white chalk and a second pattern (T2) of diagonally cut squares were used. The metal object analyzed, on the other hand, had a more uniform surface due to the texture of the material and the primer applied to protect it from corrosion. A red permanent marker was used to accentuate the color of the metal sample using the above drawing patterns.

The use of white chalk for the concrete and wood specimens, along with a red permanent marker for the metal specimen, was strategically chosen to create a high contrast between the colors of the test specimens and the artificial texture patterns. This deliberate contrast, coupled with the varied pattern representations on the object surfaces and precise exposure to auxiliary lighting, is designed to optimize the detection of a substantial number of elements by SIFT-type algorithms as discussed in Kanan and Cottrell (2012) and Wang et al. (2023).

In the SfM, the photo acquisition stage is key in producing high-quality 3D models ( Caldera-Cordero and Polo 2018). Accurate image capture is essential for ensuring precise alignment, detailed reconstruction, and reliable measurement outcomes, particularly in structural analysis and laboratory testing scenarios.

The photographic capture process for this study incorporated scale bar (SB) configurations, image overlap percentages, and camera calibration techniques to facilitate the analysis of the relevant variables. These configurations and processes align with the methodologies outlined in de Moraes and da Silva (2024).

For each analysis, images were acquired following an ordered variation in camera positioning to establish a capture method like a regular vertical grid model. Figure 5 shows the photographic capture process for the region of interest on the concrete object using only vertical captures.

A full-frame Canon EOS R camera, paired with a Canon RF 24-105 mm f/4L zoom lens, was employed for the photographic process. The camera was configured to manual mode, with a fixed focal length of 35mm to ensure consistent settings across all image captures. Exposure compensation was adjusted to +1 EV to enhance the brightness of all captures. This setup provided precise control over exposure parameters and captured regions with sufficient detail to meet the submillimeter precision requirements of the study.

To minimize image noise and ensure a greater depth of field, each shot was taken at ISO 100 and an aperture of f/11 was selected to improve focus across the scene, ensuring sharpness across the entire field of view. To further ensure the accuracy and stability of the images, a tripod, and a 5-second timer shutter were used.

A 1-meter capture distance was adopted in the experiments to evaluate the SfM technique used in structural testing. This distance maintains Ground Sample Distance (GSD) values and provides a more efficient capture process. The GSD values obtained for the experiments were approximately 0.15 mm. The image settings were configured to 6,720 x 4,480 pixels, with a 35 mm focal length on a camera equipped with a 35 mm × 24 mm full-frame sensor.

The photographic images were in RAW format. However, the image sets were converted to JPG and TIFF formats to evaluate the impact of different storage formats on the quality of 3D modeling. According to Detchev et al. (2014), Morgan et al. (2017) and Verma and Bourke (2019), these two formats are widely used in projects employing SfM. We chose the commercial software Adobe Photoshop ( www.adobe.com/br/products/photoshop.html) to convert the raw files captured by the camera to TIFF and JPG formats, because it is easy to use. However, open-source computer applications such as RawTherapee ( www.rawtherapee.com), darktable ( www.darktable.org), and GIMP ( www.gimp.org) also successfully meet our needs.

A Starrett EC799A-8 digital caliper was used to measure the lengths of the sets of SBs and CBs to size and check the 3D models developed. The instrument accurately checked the distances between the markings on the rulers and the targets of the reference elements to sub-millimeter accuracy. It offers a prominent level of precision, with error margins of ± 0.02 mm for measurements up to 10 cm and ± 0.03 mm for measurements above 10 cm Company (2007).

Five measurements were taken for each positional element to determine the length of the SBs and CBs accurately. The accuracy of the reference elements, combined with the positional accuracy of the SfM modeling, played a crucial role in determining the accuracy values used in the evaluation of the developed models. Figure 6 shows the mean values of the length of each bar obtained from the set of measurements, together with the standard deviation for each positional element. All measurements of the bars and images utilized in this study are freely accessible in the OSFHome repository at the following link: https://osf.io/k82ar/.

Initially, from a set of input images, the SfM process generates a point cloud referenced to the camera coordinate system, resulting in an inaccurate representation of real-world objects. Therefore, a scaling procedure is essential to scale the point cloud relative to a specific reference unit and ensure the positional accuracy of the three-dimensional product ( Luhmann et al. 2020). Different formats and positioning configurations of SBs with known lengths were used to scale and refine the generated products.

A series of procedures were applied to evaluate the length of the positional elements used to estimate the positional accuracy of the models. First, the lengths were obtained by measuring the generated virtual models and compared with the values obtained from the digital caliper measurements. As discussed by Garcia and Oliveira (2021), the RMSE value ( equation 1) was used to analyze the error and evaluate the accuracy of the distance prediction. RMSE = ∑ i = 1 n ( E i − R i ) 2 n (1) where n is the number of samples, E _i is the estimated value at position i, and R _i is the value measured at position I.

The analysis of the values and the quality of the adjusted data provided by each modeling, via maximum and minimum values of the covariance matrix, facilitates a comprehensive evaluation of the entire 3D reconstruction process.

Image sets with an overlap of approximately 80% were used to generate different 3D models. Camera calibration parameters were obtained using the pre-self-calibration method, using the maximum number of SBs (8) in environments like the region of interest, with consistent lighting configurations for each analysis. Figures 7 and 8 show the results of the RMSE values for these configurations in the concrete samples and the maximum and minimum diagonal values of the covariance matrix.

Consistent RMSE values ( Figure 7) between 0.12 mm and 0.11 mm were observed for the positional quality of 3D modeling for the different lighting setups in the concrete specimens. However, in terms of the quality of the modeling adjustment ( Figure 8), a significant improvement was achieved when auxiliary lighting was used in the photographic capture process compared to the natural lighting configuration (CNA). The values averaged 0.34 mm ² and 0.13 mm ² for the maximum and minimum values of the covariance matrix respectively when additional lighting was used, whereas the natural configuration gave significantly higher results.

These patterns, which indicate an improvement in RMSE values and the quality of 3D model adjustment when additional lighting was used, are attributed to the spectral characteristics of the material used as the test specimen. According to Senevirathne et al. (2021), concrete materials are influenced by factors such as mixture composition, texture, and surface color of the object, leading to a higher reflectance rate and hence more accurate 3D modeling of objects even under limited lighting conditions, as observed in the experiments.

As photographic capture relies on the amount of light reflected from objects to ensure a clearer process, the generated models showed a higher quality of adjustment under higher lighting intensities (e.g. in CNB, CNC, and CND configurations), due to the greater detail of the analyzed objects and areas of interest displayed in the image sets from these configurations, resulting in more accurate modeling.

Figures 9 and 10 show the results for RMSE and maximum and minimum values of the Covariance Matrix, respectively, for the different lighting configurations on metallic samples.

The application of a layer of protective paint to the surface of the object is common due to the characteristics of metallic specimens concerning the oxidation process and rust formation. Although the paint protects the metal from corrosion effects, as a side effect it also smoothens the surface of the analyzed object, further reducing the textural properties of the specimens, as discussed by Sudarsanan et al. (2019).

The experiments showed a slight improvement in positional quality when any form of additional lighting (MNB, MNC, and MND) was used compared to its non-use (MNA). The RMSE values for the use of additional lighting ranged from 0.19 mm to 0.25 mm and were 0.32 mm when the MNA setup was used.

Regarding the maximum and minimum values of the Covariance Matrix for the different lighting configurations of metallic objects, a significant improvement was obtained when any form of lighting assistance (MNB, MNC, and MNC) was used. The elements of the Covariance matrix showed maximum values of the order of 0.4 mm ², which is like that obtained when no additional lighting (MNA) was used. However, compared to the standard model of the test specimen, the minimum values improved by about 0.1 mm ² when any lighting aid was used.

Finally, the results of the 3D modeling of the wooden specimens provided RMSE values and maximum and minimum values of the Covariance Matrix as shown in Figures 11 and 12, respectively.

The RMSE values of the positional quality of the modeling across the different lighting arrangements in the wooden test bodies showed uniformity, ranging between 0.12 mm and 0.15 mm. Additionally, the quality of adjustments improved when any form of auxiliary lighting (WNB, WNC, and WND) was used in the photographic capture process, compared to the natural configuration (WNA).

The results of the quality analysis of the experiments with the wooden specimen can be attributed to the different structural characteristics of the materials. As discussed by Feng et al. (2019), the texture of wood shows a wide range of color variations and patterns that can coexist in a single artifact, allowing for detailed photographic capture comparable to that observed in concrete objects. The variety of texture and contrast of the objects’ surfaces facilitated a more detailed capture process, particularly when appropriate lighting was used, improving the recognition of elements in the image sets used and resulting in highly accurate modeling.

The analyses revealed the integration of lighting assistance substantially improved the quality of the 3D modeling for the three materials examined. However, the number of elements detected in each set of images, considering the various configurations and materials used, was evaluated towards a more comprehensive analysis for the selection of the most suitable configuration and, hence, optimal results. Figure 13 displays the number of sparse cloud points obtained after the detected elements had been filtered.

The number of points obtained for each lighting configuration shows a subtle consistency among the values for the same material analyzed. However, the most favorable results were obtained using the “Vertical (B)” and “Adjacent (C)” lighting configurations. A comparison of such information with previous analyses showed a slight advantage for these lighting configurations in terms of quality parameters.

Despite its results with minimal variation, the “Beneath (D)” configuration posed significant challenges to equipment installation and usage within a laboratory setting. Space constraints and safety considerations in structural testing, particularly beneath the test specimens, compromise the practicality of implementing Model D. Although the configuration offers advantages, organizations must carefully assess its adoption in terms of safety and test feasibility.

Due to the proximity of the values, auxiliary lighting should be positioned directly in front of the object or adjacent to the region of interest for photographic captures aimed at modeling objects with submillimeter precision.

New sets of capture and processing involving the application of artificial textures to the specimens were initiated as a function of previous findings that underscored the advantages of “Vertical (B)” and “Adjacent (C)” lighting aids during the photographic capture stages. Each material analyzed in the experiment, namely, concrete, metal, and wood, was examined with two distinct texture patterns (T1 and T2) and the natural pattern inherent to each specimen.

Figure 14 shows the modeling results of the concrete specimens, in which two different lighting configurations and three texture models were analyzed. The artificial texture models were crafted, in a checkered pattern using white chalk to accentuate details of the surface of the object in the first model and the pattern used in the second.

The results showed a notable equilibrium in the total RMSE values for all configurations, averaging around 0.11 mm. However, the covariance matrix values ( Figure 15) showed differences between the artificial and natural texture models, especially in the maximum values. There was an improvement of approximately 0.1mm ² when artificial texture models were used.

Figure 16 shows the results of the RMSE quality metrics for modeling metallic samples. Texture patterns were created in the analysis using red permanent markers to enhance the contrast with the color of the protective paint applied to the samples. The first artificial texture (T1) aimed to highlight detail by incorporating a checkerboard pattern on the surface. Conversely, the second model (T2) intensified the pattern introduced by T1.

The T1 texture configuration showed a slight improvement in the RMSE results, with average values of 0.16 mm compared to the other texture configurations, which achieved average values of 0.21 mm. As shown in Figure 17, there was a significant improvement in both the maximum and minimum values of the Covariance Matrix when each artificial texture model was used.

Figure 18 shows the RMSE results of wooden object modeling with configurations of artificial textures like those used in concrete specimen experiments. The artificial texture models on the wooden specimen were created using white chalk in two different patterns to accentuate the surface details of the object. The first pattern (T1) follows a checkered pattern while the second (T2) has a denser pattern.

The behavior of the wooden specimens was remarkably like that of concrete objects. Although the RMSE values did not show significant variations with the application of different texture settings, the adjustment accuracy of the 3D models improved. Figure 19 shows the maximum and minimum values of the diagonal of the Covariance Matrix for modeling the wooden specimen, emphasizing a notable improvement in quality in terms of maximum values when an artificial texture model was applied.

Analysis of the results for the three materials showed similar behavior, despite the distinctive characteristics associated with the natural texture patterns and color of each sample. The use of artificial standards resulted in more accurate 3D modeling with less variation in the maximum and minimum values, as shown by the covariance matrix values. However, no significant gains in modeling were observed when positional quality was analyzed using RMSE values.

The consistency of the positional quality values is due to the image sets used in the analysis. The use of artificial lighting in all analyzed sets resulted in highly accurate modeling, as previously investigated, and combined with artificial textures, led to the acquisition of point clouds with low representation errors, as discussed by Hafeez et al. (2018).

The analysis suggests that artificial textures tend to improve the accuracy of 3D object modeling and facilitate the detection of elements between images, resulting in a more detailed representation. The experiments involved the use of light tools and artificial textures with significant contrast to the materials analyzed. It is therefore the responsibility of the user to determine the most appropriate pattern and methods for representing the texture of an object, considering the specific requirements of their experiments in terms of feasibility and potential benefits. In engineering laboratories, particularly for structural testing requiring sub-millimeter precision, the use of artificial textures with checkerboard patterns and a high degree of repetition is recommended to improve the quality of 3D modeling.

Additional 3D modeling processes explored the effects of using different storage formats (TIFF and JPG) in sets of images captured in indoor environments at close range. Artificial textures (T1) and lighting support (B – Vertical) were adopted for the three test specimen materials.

The combinations made, including formats and various configurations, are presented in Table 2, along with the corresponding amounts of storage used by each format. The image sets indicate that the average size of files in TIFF format was approximately fifteen times larger than those in JPG format.

Figure 20 shows the quality values associated with the different formats for storage image sets for the optimal lighting and texture configurations previously examined in this study.

In comparison to JPG, the variation in the RMSE values for each specimen of the varied materials analyzed showed a slight improvement when TIFF was used. Such a trend is in line with the findings reported by Detchev et al. (2014), who observed that improvements in positional accuracy when using raw storage formats, as opposed to JPG, are typically of the order of submillimeter. Therefore, where file storage and transfer are important considerations, the adoption of formats such as TIFF may not offer substantial benefits and may even increase storage requirements.

The RMSE values did not reveal any significant differences that would justify the selection of a specific storage format. However, the results of the Covariance Matrix showed substantial disparities, as shown in Figure 21. The 3D modeling with images in TIFF format showed a better adjustment quality compared to models using JPG images.

The maximum and minimum values were notably more accurate and precise when the raw format was used. Morgan et al. (2017) reported this behavior, recognizing the higher level of detail in images in TIFF format and choosing raw formats over compressed ones. This decision was based on the superior capacity of TIFF for accommodating post-processing techniques while maintaining the integrity of the raw image data with no compression and information loss typical of JPG format.

TIFF storage format is recommended in function of its potential for higher efficiency in image post-processing techniques, especially in 3D modeling tasks requiring submillimeter precision. However, due to its rapid acquisition and lower storage space requirements, JPG can be a viable alternative when precision requirements are less stringent and extensive image post-processing activities are unnecessary.

This study, centered on the three-dimensional modeling of objects within the context of laboratory structural testing, has identified some limitations inherent to the experimental approach, particularly with SfM techniques.

Capture Distance: The experimental protocol required photographic captures to be conducted at approximately 1 meter from the objects under study. This distance, dictated by stringent safety protocols in the laboratory environment, inevitably constrained the resolution and detail of the 3D models produced. It is recognized that shorter capture distances would likely enhance model quality by increasing image resolution and enabling a more comprehensive representation of the object. Consequently, it is imperative to establish minimum capture distances that meet safety standards and optimize the quality of 3D reconstructions.

Artificial Texture Patterns: The application of artificial texture patterns in this study was intended to enhance the visibility of surface features on the test specimens, a critical factor for SfM algorithms. However, closed or overly repetitive texture patterns can obscure finer surface details, such as cracks or microfractures, which are essential for accurate structural analysis. It is therefore crucial for laboratory professionals to select texture patterns that balance enhancing surface visibility and preserving the detectability of critical surface features. This selection is vital for ensuring the robustness of the 3D models generated through computer vision techniques.

Auxiliary Lighting Positioning: The complex environment of the structural laboratory, characterized by the presence of various equipment and sensors, poses significant challenges for the effective positioning of auxiliary lighting. Inadequate lighting arrangements can introduce shadows and uneven illumination, which can degrade the quality of the images captured and, consequently, the accuracy of the 3D models produced. Proper lighting placement is essential to mitigate these effects, ensuring consistent illumination and minimizing shadowing that could compromise the integrity of the SfM process.

In summary, professionals engaged in such experimental work must possess a thorough understanding of the specific requirements and limitations of the tests being conducted, as well as the characteristics of the laboratory environment. This understanding is crucial to avoid suboptimal capture processes that could lead to reduced modeling quality or necessitate the repetition of experiments. By addressing these limitations, the fidelity and reliability of 3D models generated through SfM can be significantly improved.