Documentation techniques

Since the second half of the 1800s the stones have been the subject of academic interest, and accordingly have been documented using various techniques: sketches, rubbings, tracings and photographs. Lindqvist and his peers painted their interpretation of the carvings directly onto the stones; these photographs of painted stones have been the primary resource for academics. Based on the tracings and rubbings, some of the images of painted stones are the only remaining trace of the original carvings due to heavy weathering of the limestone. The basis of almost all modern research is based on these painted interpretations. More recent documentation using 3D scanners, RTI and photogrammetry has allowed more researchers to study the carvings digitally rather than in the field. 

Photogrammetry

The project goal of preservation was an important consideration for our choice of digitization method. The requirement to capture accurate colour data meant that the only choice was photogrammetry. Some projects choose to rely on laser or structured light scanners because they have specifications of high accuracy, despite not recording true point-for-point colour data. Optical 3D scanners also require targets to be placed on and around the object being digitized in order to scan areas larger than approximately 1sqm. 
Placing targets also add tedious work for set-up and clean-up, but, more importantly, from a conservation standpoint targets are not ideal and the potential occlusion of important details is unacceptable for digitization of cultural heritage objects. Modern photogrammetric processing software does not require targets, can generate per-pixel colour data and surface normal vectors; making it the most powerful and flexible digitization tool for cultural heritage. Photogrammetry is scale invariant and allows modelling of both tiny and massive objects. All that is required is a good camera calibration process and accurate scale reference. 

Fieldwork 

The current context of the picture-stones can be generally classified four different situations: 1) in museum storage – primarily on pallets, 2) on display in museums or in churches, 3) built into existing buildings – primarily churches but also houses and ruins, and 4) in the wilderness – in forests, fields or on private property. Accessibility is an issue and impacted the accuracy and resolution of some photogrammetry projects. For example, some stones are the bottom or top of small niches in churches; the one at Mästerby Church was so small that the camera itself could not even fit. We had to come up with creative ways to get the lights into position for optimal lighting in many situations. Some stones in fields and forests were overgrown with trees and shrubs which had to be trimmed back, so in addition to the photography gear a selection of gardening tools, saws and axes were always in the van. 

The digitization portion of the project was to be completed in the first two years: 2019 and 2020. We did not begin digitization until January 2020, and it is still ongoing in spring 2023 due to four factors: 1) initial equipment development and upgrades took many months longer than expected, 2) Covid meant working hours were reduced and accessibility was non-existent for the many stones in museum storage and in churches 3) the increase in the amount of data that had to be captured as described above and 4) the difficult accessibility of many stones meant that during a full day of fieldwork we could only digitize a few stones: a very good day was perhaps 10 stones and an average day 5 stones. On many occasions stones were impossible to locate, and thus noted as lost, or required additional equipment to access such as scaffolding, safety equipment or even keys to locked attics or cupboards. In general reflection, photogrammetry is the ideal technique for this type of project: fast, flexible, highly accurate, and easily archived for future reprocessing with better algorithms.

Equipment

To achieve accuracy comparable to other 3D scanners, we discussed several options for the photogrammetric equipment including high resolution mirrorless and DSLR cameras. Many modern cameras and lenses utilize mechanical image stabilization by directly shifting and tilting either the sensor or lens elements; this is sub-optimal for metrological quality results comparable to other 3D scanners due to the mathematics of camera calibration. 5-axis sensor stabilization across each image in a project cannot be solved with any lens model, meaning errors are forced out in the resulting model.
A metric camera is a requirement for high precision photogrammetry, and Alpa Capaul & Weber had published a paper about a metric medium format camera that was interesting, especially considering the recent launch of the 150mp Phase One IQ4 camera. According to calculations, the 150mp sensor paired with a 40mm lens at its closest focusing distance would result in a ground pixel size of 0.042mm. At an accuracy of 0.5px, the real-world accuracy would be 0.035mm at this resolution. 

The number of photos required to capture each stone would be reduced by the large field of view; a 24mp Nikon D780 and 28mm lens would require 120 images to digitize an area of 1m by 3m at the same resolution, while the custom Alpa Add|Metric can cover the same area in 14 photos. Another key benefit of medium format cameras are leaf shutter lenses which allow the use of high shutter speeds with strobes to over-power direct sunlight. 
The major disadvantage of medium format cameras for close-range photogrammetry is shallow depth of field. This is not ideal for smaller objects or round objects, but for flat surfaces like the carved surfaces of the picture-stones it is not an issue. 

Approximately 9 months into the project we had our first prototype metric camera from Alpa. The only shutter available for this configuration was the e-shutter 250, which meant we could not utilize high speed sync. The larger issue for the photogrammetric accuracy was that the 40mm lens was optimized for 1 meter, so when focused at 45cm, there were so many optical aberrations that the effective pixel accuracy was 1.5-2px. We found that the ideal focus distance was around 80cm where we could repeatably achieve 0.4-0.5px accuracy. This meant our real-world resolution was 0.075mm and volumetric accuracy approximately 0.08mm. 

The external Silex II controller did not sync very well with the IQ4, resulting in frequent black frames and lock-up. In autumn 2021 we received version II of the camera from Alpa which incorporated the new Phase One X-shutter which would eliminate the external Silex II controller and allow us to sync with our Profoto B10+ strobes up to 1/1000s. This streamlined version made fieldwork much quicker and with more ergonomic grip it was much easier to use in difficult situations. 

Photogrammetry requires scale reference either in the images or between camera pairs. We developed relatively low-cost but highly accurate scale-bars made of Invar (Nilo-36) for the project. The scale bars are L-shaped to allow two axis scaling using three holes to hold optical targets from Hubbs Machine. The targets can be replaced if damaged, and white and retroreflective targets are available in multiple sizes. 

Method

Two large Profoto stripboxes were placed on either side of the stone for smooth and even illumination. Scalebars were placed on the ground below/beside the stones, or on top depending on the situation. The images were captured as terrestrial image strips in a Zig-Zag pattern across the surface of the stones. So starting from the top left, moving to the right, and then downwards, and then across to the left and so-on. To model the edges of the stone, a series of photos were taken at varied angles. When accessible, the backsides of the stones were also captured. Sometimes they could be captured in the same sequence as the front, but in most cases a second set of images was taken with the scale-bars moved to the backside. 

Vertical photos were also taken on most stones as well for improving the self-calibration procedure. 

File types

Meshes are the most common 3D file type for local and web visualization of 3D models. Textured and rendered meshes are generally subsampled and the detail is retained only in the texture UV map. For web visualization of objects this is fine but for research of finely detailed carvings subsampled and smoothed meshes are useless. Due to the high number of polygons in each model, the raw meshes we created in this project are too large to visualize. We instead rely on the vertices to generate visualizations of the carvings.

Processing

The RAW photographs are proprietary PhaseOne files, which require conversion to two different formats for photogrammetric processing and archiving: Jpeg and DNG. Agisoft Metashape was used for processing and each model is saved as a PLY mesh.

Scaling

The scale bars are measured as follows:

Short L:    -Both lengths = 9cm (or 0.09m in Agisoft)
Long L:     -Short length = 8cm (or 0.08m in Agisoft)
        -Long length = 18cm (or 0.18m in Agisoft)

The default accuracy setting in the scale-bar panel is 0.001m, and this is sufficient for our purposes. Ideally, two scale bars are used in each model to have error control and redundancy. 

Post-processing

This mesh is so high resolution it is too much data for most computers to render in realtime, so each mesh is converted to a pointcloud in Cloudcompare to enable postprocessing. Besides the RGB textured models, we use Cloudcompare to create a “depth-map”, also sometimes known as a morphological residual model. This technique is similar to the one used by Potter and Horn, though processed solely within Cloudcompare. 

Two final pointclouds are created for the 3Dhop visualization for each stone; the RGB vertex coloured model and an optimized depth map model where the greyscale values are baked as vertex colours. 

Results

Achieving metric-grade photogrammetry results is not easy, and even with the best theoretical methods there will always be difficulties achieving ideal results in the field. 

The photogrammetric accuracy averaged 0.4px across the projects, giving us an average project accuracy of 0.06mm at a resolution of 0.075mm. When two scale bars are used, scaling the model with two scale distances in each planimetric axis, the mean scaling error across 100 different projects is 0.0195mm and the standard deviation is 0.0128.

    
How to view models on the database

Raw data

The file structure includes 4 folders:

Agisoft – The Agisoft metashape files are here including the processing report
PLY – The ultra-high (full resolution) PLY mesh is here, as well as the vertices file that is presented in the database 3D viewer.
DNG – The raw DNG images
JPG – The JPG images that were used to generate the 3D model since the DNGs are too heavy. 

How to use the downloadable 3D data

The image files can be used with any photogrammetric processing software. The most important information you need is the scale bar information stated above, but it is also stored in a readme file in each folder.