Scanning the Ruthwell Cross

Scanning

A 3D digitization of the Ruthwell Cross was planned from the very early stages of the Visionary Cross project to produce digital data that could support the study, visualization, and presentation of the artwork both as a whole and down to its fine carved details. For this reason, even though the object of interest is quite big (more than 5 meters in height), performing the acquisition using faster mid-long range scanners was not a valid option: the fine details would have been lost due to the poorer resolution and accuracy of those scanning devices (order of a few millimeters).

Because close access was necessary for this level of detail, the first thing we did for the Ruthwell Cross was hire a professional scaffolder to build a scaffold around the monument. This was to allow us to get close enough to the cross to scan it, and also to allow us access to the top of the monument. At Ruthwell, the process of building the scaffold was complicated by the fact that the church itself is a listed monument: in addition to not touching the Anglo-Saxon cross, we also had to avoid touching the later medieval church and the nineteenth-century church fittings. Moreover, even though the cross is inside a church, its placement is not easily accessible: the cross resides in a pit behind the altar, and it almost reaches the top of the apse. The pit is around 1.5 meters deep and has a metal rail all around it, which could not be removed because it is of historical interest. Therefore, it was necessary to create multi-layer scaffolding to work at three different heights. The Cross was acquired in three days of work, starting from the top layer and dismantling a portion of the scaffolding every day.

Complicating matters was the issue of security: the Ruthwell cross is easily accessible and frequently unattended; once our scaffolding was in place, it would be difficult to prevent members of the public from using this to gain unauthorised access to the top of the monument when our team was not on site, placing themselves and the monument at risk of physical harm. This meant that the scaffolding needed to be constructed with as little lead time as possible before our team arrived to begin the scanning. It also needed to be broken down as soon as possible after we were finished.

Once the scaffolding was in place and the research team assembled from Pisa, Lethbridge, Pennsylvania, and Leeds, the scanning and photographic work began. We decided to use a laser triangulation scanner (Konica Minolta Vivid 910) to ensure sufficient sampling density and accuracy in the acquisition of all the required data. The scanning campaign, which took place between the 18 and 20 of April 2012, was planned carefully: the necessity to place the scanner at a short distance from the object (50-130 cm) was one of the reasons the scaffolding was essential.

During the scanning phase, to maintain high data quality, it is essential to control two factors: lighting and complete coverage of the surface. Given the nature of the object, it was useful to supplement the scanning phase with the Easy3D Scan tool, produced by the Visual Computing Lab. This programme allows real-time monitoring of which areas of the object have already been measured, thereby preventing ‘gaps’ in the model. No special precautions were required regarding lighting. The scan was carried out in an enclosed environment, while the light filtering through the small skylights in the ceiling proved to have no effect.

Processing

3D scanning is actually more like radar than photography. It works by projecting a signal, in this case a laser, on an object and then using the reflection to calculate the location of individual points. The resolution of the scanner determines how close the points measured can be. In our case, we did most of our scanning at a resolution of 0.25 mm, meaning that we had four measurements per millimetre of surface on the cross. A total of 487 range maps were obtained, plus a further 23 acquired from a stone fragment believed to be one of the missing parts of the cross, each with an error estimated at approximately 0.2 mm.

The raw data were processed following the usual 3D scanning pipeline, using the MeshLab tool [Cignoni et al. 2008]. It was used for semi-automatic alignment, manually identifying four or five points shared by the two range maps, and then applying the ICP algorithm generally every five rough alignments. To prevent the calculations from becoming too computationally intensive, the various alignments were divided into 20 groups, each containing a minimum of five range maps and a maximum of 40. Dividing the cross into vertical sections served not only to keep the calculations manageable but also to prevent an accumulation of error that would misalign the object longitudinally. At this stage, data quality was preserved thanks to the large number of scans taken and the generally extensive overlap between adjacent scans. The various areas of the surface were found to be rich in data, minimising the misalignment of scans prior to the merging phase. At the end of the alignment phase, the average error found in the data was approximately 0.5 mm. The noise present was minimal, and the filtering stage involved practically only the removal of the scaffolding sections that had been captured.

For the Ruthwell Cross model, 373 range maps were merged using MeshLab’s volumetric filter known as VCG. A resolution of 0.7 millimetres was selected, resulting in a 3D model comprising 112 million polygons. The quality of the result clearly depends largely on the results of the previous stages. The algorithm used in the merging stage was one of the best available in 2014, and the resolution, considering the number of triangles, is decidedly high, allowing for an excellent level of detail and limiting data loss to imperceptible levels. For example, starting with scans with an acquisition error of 0.2 mm, an alignment was achieved with an average error of 0.5 mm. Finally, the merge was performed at a resolution of 0.7 mm. This indicates that the model is capable of representing details as small as 0.7 mm; this information is crucial for the study and analysis of the geometry, as experts can ascertain the accuracy of every measurement taken on the 3D model.

Colour

The colour of the Ruthwell Cross was captured using standard photographs. Over a hundred photographs were taken during the scanning process. Unfortunately, several problems arose during the alignment phase, as the photographs turned out to be too distorted. In any case, this initial alignment phase, which Federico Nomellini was tasked with, produced 78 alignments. It was decided to take a further series of photographs to improve the colour quality. The new series comprised 259 shots, from which approximately 200 usable alignments were produced.

The alignment procedure in MeshLab combines a manual component with an automatic one. You begin by loading the various images as layers, much the same way as in photo-editing programmes. The programme allows for intermediate views between the photo and the model, with the option to increase or decrease the model’s transparency. The operator must then move the model into a position as similar as possible to that in which it is depicted in the photo; at this point, using the MutualInfo filter (SCOPIGNO et al., 2009), the programme processes the data to refine the alignment. Furthermore, in order to reduce the interference caused by optical distortion that may be present in some photos, it is possible to vary the field of view (FOV). As the name suggests, the filter uses a system based on mutual information that identifies similarities between two images—in our case, the photograph and a rendering of the model. The algorithm adjusts the camera parameters until it finds a position that maximises the mutual information between the image and the 3D model rendering.

The accuracy of the data at this stage was ensured by the large number of images of the Cross used, the overall quality of those actually used for the alignments, particularly with regard to distortion, and the precision of the alignment algorithm. Once the images had been aligned to the model, colour projection was achieved using the masked photoblending method (SCOPIGNO et al., 2008), which assigns a colour value to each portion of the surface that takes into account the contributions of all the images projected onto that portion. Unlike the geometry acquisition procedure, colour projection does not allow for much control over the data. This is due to the input data, since the source images are not calibrated, and the lack of an objective measure of the quality of the projected colour. The ability to project colour onto the surface allows us to create a model that we might describe as more realistic, even though this is a component that affects only its appearance; however, in some cases this appearance is useful in enabling a better understanding of the object’s details.

Results

The final “master model”, reconstructed from the sampled data with a cell resolution of 1 mm, is a heavy 112 million triangle model. Two further models were created at lower resolutions, 15 million and 5 million triangles respectively, using another MeshLab tool, namely MeshSimplify.

The entire digitization activity took four days of work on site (shape and color sampling), while a few weeks were needed for post-processing the raw data to produce several models, with different resolution and color encoding modalities, required for visualizing the Cross on different hardware platforms. Moreover, several higher-resolution models of portions of the Cross surface were also created to account for more focused analysis. This resulted in a photorealistic 3D model of the Ruthwell Cross that can then be manipulated in a variety of ways to emphasise different features: adjust or remove colour and other texture, artificially create shadows, enlarge or shrink the resolution, or even send the object to a 3D printer or router to manufacture highly accurate physical replicas.