Rock Art in Three Dimensions: Comments on the Use and Possibilities of 3D Rock Art Documentation

Abstract

Methods to document rock art in all three dimensions have become a standardized workflow. In this article, we discuss their advantages and disadvantages when compared to older reductive approaches to rock art documentation. Furthermore, some misunderstandings regarding 3D documentation are addressed. As the majority of the problems presented by the 3D documentation of rock art can be solved through advanced visualization workflows, recent developments in this area are described. The rock art documentation described in this contribution also serves wider research purposes, which will be discussed. Newly discovered images and newly developed machine learning algorithms will also be introduced.

1 Introduction

The strong push to make image and range-based 3D recording the standard methods of documenting rock art has largely been successful in hindsight. The best evidence for the overall scientific acceptance has been the successful granting of several large digitization projects concerned with rock art in Italy, Great Britain, Iberia, and Scandinavia. The material that is digitized is diverse in technique and chronology, dating from the Neolithic and the Bronze Age to medieval picture stones and later expressions of human creativity on rock (Alexander et al. 2015; Díaz-Guardamino et al. 2019; Díaz-Guardamino Uribe and Wheatley 2013; Foster et al. 2016; Horn and Potter 2019; Höll et al. 2014; Noble and Brophy 2011; Oehrl 2020). While this seems like a success story, reluctance and misconceptions still surround the 3D documentation of rock art. However, proponents of 3D documentation should be open to concerns about problematic aspects of these methods. This is not only an issue of academic ethics, but also that it helps us to address these problems and advance our methods. Baseline method development is often treated as somewhat of a black sheep by funders and researchers who are both often just looking forward to the next earth-shattering discovery rather than the baseline work that enables such discoveries. This neglect is the reason why the “third science revolution” in archaeology was driven by fields outside of archaeology and had very little initial involvement from archaeology (Kristiansen 2014). It is also the reason why many archaeologies and archaeologists are just now catching up with debates that should have been present at the onset of the developing methods – as best illustrated by the contentious discussion around aDNA (Frieman and Hofmann 2019; Furholt 2019; Pääbo et al. 2004).

In the following, we will discuss the advantages of 3D documentation compared to reductive, traditional methods by reviewing some of the misconceptions. We will also discuss the problems that exist with 3D documentation, and finally, outline some future directions regarding how 3D recordings of rock art could be used to accommodate some of the problems and drive forward future research. The following remarks will only address rock art during the Nordic Bronze Age (1800/1700–550 BC) which was produced by a percussive technique removing material from the rock surface to produce a negative relief (Fig. 6.1). Although not technically correct in most cases, the images are called carvings. We will keep this term and use it interchangeably with the term petroglyph. We will give the inventory numbers of the Swedish National Heritage Board for each rock art panel we discuss.

Fig. 6.1

Rock art distribution in southern Scandinavia (green square: main study area)

2 2D and 3D Rock Art Recording

In the best of cases, petroglyphs are several millimetres deep and are visually perceptible. However, most carvings are very shallow either through weathering and other erosion (Robinson and Williams 1994; Sjöberg 1994; Swantesson 2005), or because they were initially not carved very deeply. The latter may seem counter-intuitive, because why would people have made images which they then could not clearly see. Freshly produced rock art removes part of the weathered rock surface which makes it show up lighter than the surrounding areas. This means that the images would have been visible to observers for at least a few years even if they were very shallow. A carving in Lyse (L1969:9961; Lyse 142:1) was discovered on a vertical surface protected from weathering by an overhang, which still has some of its lighter appearance preserved. Experiments in carving rock art (Bengtsson 2004; Lødøen 2015) and even modern vandalism show that rock carvings are lighter than the surrounding rock (Fig. 6.2). Due to the problematic visibility and continuous erosion of rock art, documentation methods have always strived to be as holistic and unbiased as possible (Horn et al. 2018; Nordbladh 1981).

Fig. 6.2

The large human figure on a panel in Aspeberget (RAÄ Tanum 18:1; L1967:2415) is an example of modern vandalism, but also shows how freshly carved rock art is lighter than the surrounding rock. The strongly white figures are the result of painting in Bronze Age rock art for the purpose of documentation. (Image: Catarina Bertilsson, SHFA (CC-BY NC))

Even though they could not record it, the depth of the petroglyphs was always a crucial feature even for traditional methods, i.e. rubbings and tracings. These methods included several interpretative steps starting from a tactile survey of the rock’s surface to feeling differences in the grain structure that would enable the investigator to differentiate between carvings, natural features, erosion, and damage (Milstreu and Prøhl 2020). At best, several documenters are involved, mutually checking their interpretations (Most recently Toreld and Andersson 2015). The result of this investigation is then transferred to the rock using chalk, which is then transcribed to plastic sheets using drawing or stippling techniques. Alternatively, paper is put over the area which is then rubbed with graphite (summarized in Horn et al. 2018; Nordbladh 1981).

Each of these steps includes considerable human bias, because the decision of whether a line is carved or a natural feature relies heavily on experience, expectation, and assumption (Bertilsson et al. 2017; Horn et al. 2018). This bias is directly inscribed into the fabric of the documentation, which means reductive documentation is always already interpretive. This interpretation is then carried over into the investigation of research questions perhaps by other researchers, who may find it difficult to trace this back. The second major disadvantage of traditional methods is their reduction of the three-dimensional heritage into a two-dimensional plane. Although some cues to the depth differences in carved lines can be gleaned from rubbings, and some newer tracings which attempt to indicate superimpositions (for example Toreld and Andersson 2018), this is yet another layer of interpretation because tracings and rubbings both do not depend on real numerical depth values (Horn et al. 2019). Tracings provide an additional major problem since they are an abstracted representation of the surface structure, for example erosion patches are shown in red circles and ice-lines are only shown by a single rough indication of their overall direction. This is problematic, because natural features are sometimes included in the carvings and would then be erased from the documentation (Fig. 6.3a–b). There are several additional minor problems, but they have been described in detail elsewhere (Bertilsson et al. 2017; Horn et al. 2018). Despite all the criticism, it has to be acknowledged that the reductive methods provide compelling, easy to read, and clear depictions, which is perhaps the reason why they are still preferred in the illustration of research articles.

Fig. 6.3

Three documentations of a panel in Brastad (L1970:9162; 142:1) demonstrate the inscribed biases in traditional documentation (top center: tracing by Stiftelsenfördokumentation av. Bohuslänshällristningar, bottom left: rubbing by Dietrich Evers, and bottom right: ratopoviz visualization; laser scan by Henrik Zedig; visualization by Oscar Ivarsson & Christian Horn): (a) tracing shows a continuous line as interrupted, but discards the fissure which goes through the line; (b) the rubbing focusses on the hand and almost omits the stronger object that superimposes it; (c) the rubbing presents the hour-glass figure too weakly and the tracing omits the intersection with the human; (d) the tracing omits details of the right prow of the boat and the rubbing deemphasizes the top part of the prow. (All images provided by SHFA)

The 3D recording methods used most in rock art documentation are laser scanning and Structure from Motion (SfM) photogrammetry. Another photogrammetric approach that provides promising results is Reflectance Transformation Imaging (Bertilsson et al. 2014, 2017; Bertilsson 2015; Horn et al. 2018; Horn et al. 2019; Meijer 2016; Meijer and Dodd 2018, 2020). Newer proposed extensions to these methods include using macro- and micro-photography for photogrammetry (Plisson and Zotkina 2015).

Reductive and 3D documentation methods have a selection bias, in the sense that a decision is made on what to record. However, baring technical issues in the recording or processing of the data, these methods simply record everything within their scope without additional bias. Thus, eliminating the recording bias of the reductive methods, meaning that the documentation itself is not an interpretation in which bias is inscribed. This allows 3D documentation to postpone the introduction of bias to the final data interpretation step which makes any scientific work more controllable. Furthermore, by its very nature 3D recordings provide a more faithful reproduction of the rock art since it has no dimensional reduction.

3 Advantages of 3D Documentation Compared to 2D Recordings

One of the advantages of 3D documentation is the flexibility in its analysis. It can be turned and scaled without any barrier, we can go from a birds-eye view, potentially even looking at multiple sites in their landscapes, to a close-up focussing on a particular image. This close-up can go into the microscopic scale assuming the right equipment like macro- or micro-lenses was used during recording. The freedom to move the viewing angle means that we can analyse several perspectives on the rocks and the images on them, and even compare several perspectives directly. This is important because documentation of rock art is usually a birds-eye overview which obscures how the images interplay with the topography of the surface (Helskog and Høgtun 2004). Multiplying these perspectives in drawings is difficult and time consuming. From a 3D file they can be produced in seconds.

In a recent publication of a reductive documentation of the famous Vitlycke panel (RAÄ Tanum 1:1; L1968:7678), the authors claim to have the advantage of viewing the panel from limitless perspectives and angles when they analyse the panel before and during their documentation (Toreld and Andersson 2018). This, of course, is true to an extent because while on the panel they can stand up, move close, and shift their perspective. With 3D documentation however, the advantage is everyone can do this after the recording. After documentation with tactile methods, the documentation is simply flat and fixed to the perspective the documenter has chosen. However, even while in the field, the documenter is more limited than Toreld and Andersson (2018) make out. True bird’s eye views, perhaps even with rapidly shifting perspectives, are almost impossible in the field. Close up observations are also limited, since to our knowledge rock art documenters rarely bring lenses to the field, and there is a strong emphasis on using the sense of touch rather than visuals.

In a recent comparison of and criticism on traditional and 3D methods, both, Meijer and Dodd (2018) maintain that rubbings do not demand interpretation. On a theoretical level, this is true. Imagine that we were somehow able to let a machine make a rubbing that exerts equal pressure and consistent graphite application across the entire documented surface. That would be a visualization that is – as Meijer and Dodd (2018) put it – purely surface based. However, in reality, this is rarely the case because documenters have experience and expectations. Known figures tend to be brought into the centre of the paper that is fixed on the rock, because the rubbing tends to be weaker towards the sheet border. This can make figures, especially those that were not known about previously, hard to spot (Figs. 6.3b and 6.4). Some figures show stronger application of graphite than others. The reasons for this may be varied, for example, someone might get excited by a figure showing up strongly, while they are less vigorous in areas where they expect nothing. Willingly or not, this is the inscription of bias into the documentation.

Fig. 6.4

Rubbings clearly showing the borders between the different paper sheets (a) and a stronger rubbing where images were suspected (b). (Rubbings by the Rock Care Project, provided by SHFA)

We are sympathetic to the demand of using multiple documentations and documentation methods for rock art research (see Horn et al. 2018; Meijer and Dodd 2018). However, we do not share their darker vision that the digital methods will remove rock art researchers from the original surfaces because it will always be necessary to engage with the original bedrock surface analysing it to find panels which merit documentation through laser scanning or SfM. It is unfeasible to record every single exposed rock art surface on the off chance that there may be something carved. Inevitably, there will be the odd misplaced target point, however, it would be a tall order to maintain that traditional techniques are free from similar human failures. Meijer and Dodd (2018) suggest that there will always be people that take whatever comes out of a computer at face value. However, we do not think that this is a change to the status quo. Secondary users of rock art documentations have mostly assumed that the documentations are for the most part reliable and – outside dedicated rock art research – have rarely criticised or questioned the method or content of any documentation directly. This is precisely the reason why we should strive for bias reduction while acknowledging that preventing any bias might be impossible.

The same authors state in a recent publication of 2D and 3D documentations from various sites in Tanum (Sweden) that rubbings are highly accurate, because “a human hair trapped between the [rubbing] paper and the surface is visible on the rubbing” (Meijer and Dodd 2020). Human vision or in this case visual acuity is a complex issue so only a few comments can be made here. Human acuity is approx. 1 min of arc which means that at the minimum focus distance of ca. 10 cm for a perfectly healthy human, the visibility limit is 0.03 mm (Diniz et al. 2018). Since normal European hair is on average between 0.06 and 0.08 mm in diameter (Jackson et al. 1972), the hair is actually very large and unsurprisingly can be seen from quite a distance away. We are, of course, speaking about theoretical possibilities, and here is where the potential of the 3D documentation methods shines. The CreaformHandyscan 700 currently employed in the work of the Swedish Rock Art Research Archives (SHFA) has an accuracy limit of 0.03: the limit of human naked eye vision. However, on screen, we can move much closer because this is not affected by human focus distance. Of course, rubbings would not pick the hair up, was it in a carved line. For 3D models, this is not an issue and with micro-photogrammetry, we can already go far beyond the capabilities of human vision (Plisson and Zotkina 2015).

4 Problems in 3D Documentation

To achieve bias reduction, and to identify development potential for documentation methods, it is important to be transparent about issues related to 3D documentation of rock art. Some of this has already been discussed, for example, that target points can cover up carved areas (Meijer and Dodd 2018). There are other procedural problems, for example, especially with larger panels where there can be gaps in the photo coverage which will only be obvious once the model has been calculated upon returning from fieldwork. Too much light is problematic for the laser scanner, and SfM can struggle with intermittent dark and light patches. Rain or wet surfaces are detrimental to both methods as the surface reflections prevent effective capture.

Bertin (1983) considers in his ‘image theory’ the capacity of each human eye to only perceive a 2D image. For the perception of three dimensions, human vision requires multiple visual cues (Welchman et al. 2005): Stereoscopic vision, accommodation, parallax, size familiarity, and aerial perspective. In the natural world, this is not an issue as it is three dimensional and provides these cues. To perceive the three dimensions of a model on a screen, either the model has to be moved or light needs to be moved across their surface to provide the necessary cues. In a still image, some of the observations based on the surface shape tend to disappear depending on the lighting angle (Horn et al. 2019).

This becomes especially problematic when we consider that 3D models easily exceed more than one or even several gigabytes in size (Table 6.1). In Scandinavia, this becomes a more pressing issue because full site 3D documentations of larger rock art sites are still underrepresented. Large file sizes can of course be made smaller, but this directly converts into a reduction of the resolution, and therefore, the quality of the model. Such steps need to be undertaken to upload models to public repositories such as Sketchfab which allows institutional users an upload size of only 200 MB. Storage on servers and in clouds is still expensive, so journals usually do not want to host multiple, large models.

Table 6.1 File size comparison

For all these reasons it is imperative to develop efficient visualization methods of 3D documentation. The aim is to show the content of the panel in a way that conveys a maximum of information including images, superimpositions, and topography, but also achieves the clear and compelling nature of reductive documentation methods. This should be achieved with as little manipulation of the data as possible to keep the information suitable for scientific inquiries.

5 Revealing Rock Art

To tackle the challenge of providing good visualizations that keep as much information as possible, we have developed three different techniques based on roughly similar principles. All the approaches that we developed focussed on curvature, depth maps, and on previous work concerned with this. The open source software Visualization ToolKit (VTK) and Paraview have been used to produce visualizations from the variation of relative and absolute elevation (Trinks et al. 2005). These produced good visualizations revealing especially deeper carvings such as cupmarks and outlined a way forward for other workflows. Another approach called “AsTrend” produces excellent visualizations. However, it uses the LiDAR Visualization Toolbox which causes distortions on slopes (Hesse 2010) and by now it is outdated receiving its last update in 2014. The authors also manipulate the relative values of pixels, for example, through the Adobe Photoshop dodge and burn tools (Carrero-Pazos et al. 2016; Carrero-Pazos et al. 2018) which meant that they cannot be used any longer to estimate depth differences. There are some other approaches which create reasonable visualizations, but that are relatively inflexible (Mark 2017). This is the background for our own efforts to create visualization methods with the aim of providing better clarity, flexibility, and applicability.

From the results of the 3D-Pitoti project, we adapted the idea to work with depth maps generated from the 3D data (Zeppelzauer et al. 2016). Depth maps are 2.5D raster images (GEOTIFF) that store depth values in each pixel represented through a grayscale colour ramp. However, visualizing rock art directly using these is problematic as the rock art is represented by height differences that are more subtle than the extreme values of the natural global curvature. On a different spatial scale, this is surprisingly similar to the problems researchers that use LiDAR data face, because the cultural remains they want to study are often obscured by larger natural height differences (Bennett et al. 2011; Crutchley 2010).

Software solutions like ArcGIS include local relief modelling tools (LRM) like Focal Statistics to deal with the issues represented by LiDAR data that have successfully been implemented to visualize rock art as described elsewhere (Horn et al. 2019). Essentially, the global curvature is calculated by blurring the depth map and then subtracting that output from the original depth map. What remains is a depth map of the local depth differences (Fig. 6.5). The result provides the same clarity as good rubbings, but has the additional advantage that the colouration is based on real depth values with a consistent spread. This means that the human bias in colour application to traditional rubbings has been alleviated, and that superimposition and depth differences can be observed in detail. Another advantage is the freedom to adjust the colour ramp and the zoom applied. Apart from 2D outputs, the LRM tools also allow images with a 3D feel by calculating the percentile used in the focal statistics step (Horn et al. 2019). Furthermore, it is possible to produce overview and detail images from the same file.

Fig. 6.5

Left part of a boat in Lövåsen: (a) rubbing (Tanums Hällristningsmuseum Underslös), (b) visualization using local relief modelling (Christian Horn, Rich Potter, and Derek Pitman), and (c) ratopoviz (scan by Henrik Zedik, visualization by Oscar Ivarsson and Christian Horn). (Provided by SHFA)

We also developed an approach that works independently of GIS software by programming an application named “ratopoviz” (rock art topographical visualization). It was specifically set up to accept common laser scan mesh formats to produce several visualizations that were subsequently used for machine learning training. Ratopoviz is an automated pipeline that samples a point cloud from the mesh using the vertices, and then removes outliers through noise detection and a clustering algorithm. The point cloud is then projected into two dimensions using Principal Component Analysis, the pixels are projected into a barycentric coordinate system and the z-values are stored in the pixels. In addition, a normal map is calculated. The global curvature is isolated with Gaussian blurring using a parameter related to the standard deviation (Zeppelzauer et al. 2015) and the removal is guided by further mathematical modelling. Through colour ramp enhancement, both in RGB and B/W using the standard deviation and blending with the normal map, six different outputs are exported as PNG files.

We have also begun developing a method to create low poly models in Agisoft (500 polys) and create relaxed UV maps (2D images representative of the 3D surface) for them in RizomUV. The 3D models are then imported into Substance Painter and processed using the high poly version of the model to bake normal maps and curvature maps. The curvature map is then enhanced using a “levels” node, which is akin to changing the contrast of the image. Additionally, a filter driven by the curvature map fills areas that feature curves. Together these create a lightweight model that highlights the areas of the surface with local differences in height and curvature while ignoring the global curvature of the surface. This allows us to generate maps and visualise entire surfaces at once and enables us to apply different colours to depth differences on a sliding scale. Altogether, with the model and 2 k texture maps the output is often less than 2 MB. These lightweight models ideal for mobile applications visualizing the rock art heritage for visitors.

6 3D Documentation and Visualization – What Is it Good For?

6.1 Preservation

To monitor rock art and help its preservation, it is important to know the exact current state of the carvings. This can be shown on some images in Lövåsen (RAÄ L1967:2412; Tanum 321:1) which the SHFA documented in the summer of 2020. On the panel there is a boat that contains a partial human figure. The head and the right arm of this figure are partially destroyed by a modern inscription that reads “1911” (Fig. 6.5). It can be assumed that the inscription dates itself. However, in a rubbing produced by Tanums Hällristningsmuseum Underslös (THU) in 2010 this is hardly visible and the number itself cannot be identified. Interestingly, in a laser scan conducted by THU on the same site, the “1911” is clearly visible (Milstreu and Prøhl 2020). Now that we know the modern damage to this rock carving and its precise form, it is possible to monitor the site for further potential and new damage. That such damage is not just a nightmare archaeologists conjure up was demonstrated by the destruction of the famous skier in Tro (Nordland, Norway) (Orange 2016).

6.2 Discovery

For research and its dissemination, it is important that the documentation conveys as much information as possible. It is in a sense surprising to find that the rubbing of the boat in Lövåsen published by THU reveals more information about the left terminus of the boat than the published snapshot of their laser scan (Milstreu and Prøhl 2020). In the rubbing, it appears that this part has a carved half oval area that accentuates the prow. This cannot be observed in the scan (Milstreu and Prøhl 2020). This possibly has to do with the light angle chosen for the screenshot illuminating the area in a way that hides this information. A visualization generated with ratopoviz using this scan shows that there are two half ovals next to each other that have only been lightly engraved (Fig. 6.5). Thus, it appears that this boat has two phases and at some point may have looked like a mix between a boat with a hull that was fully carved out and one with uncarved segments separated by carved lines along the hull. The segments were potentially applied after the hull was initially fully carved, but lightly hammered out, because they appear to be deeper.

6.3 Outreach

On the same site, it was also possible to observe how deceiving the red paint, which is applied to the carvings to guide tourists in seeing the images, can be. While the impetus is laudable, it is unfortunately often blatantly wrong meaning that tourists are misled. On panel L1967:2632 (Tanum 325:1), a spearman is painted with a short neck and a small head, seemingly stabbing a small boat (Fig. 6.6a). The 3D visualization shows a much more complex figure with a long neck extending above the spear shaft. The head is articulated and shows features that could be a combination of any two of the following: a nose, a neck ring, or a helmet. The figure has hands extending ever so slightly above the spear shaft. The lower legs and feet were either carved deeper initially to emphasize them, or were applied later. Alternatively, they could have been recarved on several occasions (Hauptman Wahlgren 2004). The supposed boat is in fact more likely an animal as indicated by the outward bend on the top, although it may have an exceptionally long snout (Fig. 6.6b).

Fig. 6.6

Spearman and boat in Lövåsen: Photo showing how it is presented to tourists in modern red paint (Rich Potter) (a), visualization using local relief modelling (Christian Horn, Rich Potter, and Derek Pitman) (b), visualisation using the enhanced curvature map (Rich Potter) (c)

Above the figure is a boat and above that a square carving. Here the carvers seem to have used a natural line to indicate what may be the handle of a hammer (Fig. 6.6b). The left keel extension of the boat seems to have been extended somewhat after the initial carving. Above the sickle-like feature is another animal which is slightly damaged in the neck area, but may be rather like the animal in front of the spear (Fig. 6.6b). Lastly, above the right prow, there is a small human figure perhaps carrying a sword (Fig. 6.6b). The keel extension, the animal, and the human were also missed in the rubbing documentation, because while graphite application was done vigorously on the boat to see all the details, the rubbing above the boat was conducted with less pressure or repetition (Fig. 6.6a). This means this area is lighter and the weaker images and details do not show in the final documentation (Milstreu and Prøhl 2020). Here we can see directly how experience and expectation imbue human bias directly into rock art documentation. Such documentations potentially inform the red painting for the tourists and present a skewed image to the public.

6.4 Going Back to the Rocks

For preservation, research, and outreach, highly detailed 3D documentation and visualization should be the standard for documenting every rock art image. Long-term storage in an online, open access repository like the SHFA, and in other places for data redundancy, guarantees the availability, viability, and sustainability of this documentation. It also means that funding opportunities for such infrastructures should exist that allow their long-term survival as otherwise this wealth of information will become more difficult to access or vanish completely. However, checks and balances are also imperative. This means 3D documentations need to be checked against documentations made with reductive methods, especially rubbings and, of course, those documenting the rocks should literally not lose touch with the rocks (Meijer and Dodd 2018).

Moreover, an overseen aspect of documentation is geological expertise, which should, if possible, be included in the process of analysing rock art sites. Work on the impact of producing rock art on the rock grain structure and its significance for the interpretation of images and engraving techniques has only recently begun, for example in the project “Tracing carvers on the rocks” (Swedish Research Council). The recent documentation of a small boat at Bro utmark (L1967:2645; Taum 192:1) whose centre is impacted by circular exfoliation indicates that some of the former image may still be preserved in the damaged area (Milstreu and Prøhl 2020). This was confirmed with a 3D recording and visualization of the boat in question. A geological analysis using a magnifying lens and microscopic camera indicated that very faint traces of the carved area were still preserved. Based on this, we were able to suggest an interpretation of how the boat may have looked originally (Fig. 6.7a–b).

Fig. 6.7

Small boat in Bro utmark: Visualization using local relief modelling (Christian Horn, Rich Potter, and Derek Pitman) (a), interpretation based on the field observations, microscopic analysis, the 3D model, and the visualization using local relief modelling (b)

6.5 Going Forwards

Some years ago, Kristian Kristiansen suggested that archaeology was undergoing a “third science revolution” of which big data was an aspect (Kristiansen 2014). The challenge is to use new methods to address the mass of data that is represented in rock art. Digital archaeology, like the digital humanities, has undergone a computational turn (Berry 2011) and machine learning has been introduced as one method that is gaining momentum. These computational methods can be used to create automatic and semi-automatic approaches to rock art segmentation, image classification, and object localization (Cai 2011; Poier et al. 2016, 2017; Seidl 2016; Zeppelzauer et al. 2015, 2016; Zeppelzauer and Seidl 2015). If such algorithms are trained on larger datasets, then they have an increased likelihood of providing a high-powered statistical tool to sequence rock art images, and discover regional groups of similar images and chronology by employing different lines of evidence. These are quantitative studies, but they are not contradictory to qualitative approaches, both should be brought into dialogue and can inform each other. This conforms with Ezra Zubrow’s (2006) theoretical outline of digital archaeology which emphasizes the very small and the very large, i.e. larger processes and detail observations.

To achieve this, initial training of a Faster R-CNN object detector (Ren et al. 2016) was initiated using the ratopoviz visualizations; this was called RAOD (rock art object detection). The training used a supervised approach with manually drawn bounding box labels on 408 laser scans and employed data augmentation and transfer learning to optimize training with the data available to prevent overfitting on the training data. RAOD identifies and localises rock art within an image and then the detection is assigned a pre-defined class label. The mean average precision (mAP) reached 32.5 for the best performing model. The class boat, with a mAP of 64, outperformed other classes while the class animal only had a mAP of 10 (Table 6.2). Boats are, despite some variation, a group with a relatively concise shape and frequent occurrences. Conversely, animals were a remarkably diverse group, even when only four-legged animals, like deer, horses, boars, dogs, etc. were included. Additionally, there were fewer examples of animals than boats available for the training dataset (Fig. 6.8). These metrics indicate that better results can be expected with additional training data and class labels to limit the in-class variation between rock art samples.

Table 6.2 Mean average precision (mAP) for several models with different input data

Fig. 6.8

Bar chart of the class labels

These outcomes of RAOD are opportunities for a dialogue to understand why the algorithm made certain decisions and to see how human and machine recognition differed. Maybe unsurprisingly, RAOD struggled with human creativity expressed in hybrid forms (Ahlqvist and Vandkilde 2018) like boats with horse heads or animals that are similar in form to boats, like the previously discussed figure on the panel in Lövåsen (Fig. 6.9a). However, the object detection results have led to new discoveries. On a panel, RAOD identified a human with 96% confidence that was neither discovered by Åke Fredsjö (1981) nor in an inspection prior to starting the training. This led to a re-inspection of the scan and the visualization and a confirmation that this human figure was present (Fig. 6.9b–c). Here, the machine assists in human-led interpretation, highlighting potential regions for rock art within an image.

Fig. 6.9

Two boats with animal head prows which the algorithm had difficulties identifying as boats or animals (Christian Horn, Rich Potter, and Derek Pitman) (a); ratopoviz output of a detail on a panel in Kville (L1969:5952; 149:1) with a red square indicating the area in which the algorithm identified a human figure with 96% confidence (b), and an interpretation of the figure (c)

7 Conclusion

The power of laser scanning and SfM photogrammetry to record the third dimension of rock art alone should warrant seeing them as the premier method of documenting rock art. As we have shown, 3D data has tremendous potential after its original capture. The potential for visualization at the very least matches the clarity of traditional rubbings, while being even more accurate in spatial relations and depth representation. Furthermore, overview and detail images even going into micro-detail given the right lens can be provided. With that, 3D documentation is an excellent source for the preservation of rock art, research, and outreach, potentially enabling us to discontinue the practice of painting the images onto the rocks soon. Of course, we should not lose touch with the rocks. As has been published earlier (Horn et al. 2018), 3D documentation should not spell the end of reductive methods, especially night photography and rubbings, because they provide important comparative material.

Beginning to train object detection and other algorithms on highly precise 3D data has proven fruitful. In the future, these first results can be used in approaches employing transfer learning and supplying additional traditional documentation data to establish new statistical tools for the research of rock art. As algorithms will undoubtedly become more sophisticated, with more powerful computers, and with more data, the future of rock art research looks bright, with new opportunities and discoveries. It is important to remember that these are tools that deliver outputs which are always in need of interpretation to give them meaning. In that sense 3D, big data, or digital archaeology are theory agnostic and not a contradiction to post-processualism or any other modern theoretical approach.