SCHEMATA: 3D Classification and Categorization of Ancient

Terracotta Figurines

Alexander Zeckey and Martin Langner

Institute for Digital Humanities, Georg August University Göttingen, Heinrich-Düker-Weg 12, Göttingen, Germany

Keywords: Object Mining, Image Exploration, Greek Terracotta Figurines, Machine Learning, Typology.

Abstract: The goal of the starting case-study is not only to develop procedures for automatically generating corpora

using 3D pattern recognition, but also to reflect on the associated schematizations and how they can be applied

in computer science and visual sciences. For this purpose, methods of object mining in 3D data are to be

developed. We chose an object group which is defined by its complexity in shape and the similarity between

the objects: In 4th and 3rd century BC ancient Greece small terracotta figurines used to be an art form that

was quite common. Based on 200 of those terracottas, a classification system will be elaborated with digital

methods, which is able to meet the complexity of the artefacts. In close cooperation between computer science

and archaeology, this experimental process leads to a fundamental examination of the concept of pattern

recognition as a humanities category. The discussion of the various concepts and methods will be carried out

in two complementary dissertations.

1 INTRODUCTION

Three-dimensional objects with complex forms are

inadequately classified both in applied computer

science and disciplines dealing with material artifacts.

Archaeologists are confronted with the problem that

although resemblance in shape can be recognized and

established, it is much harder to support it with

reasons and to describe adequately in language what

may be visible for the eye. Furthermore,

archaeologists have yet to make sufficient use of

automated 3D shape recognition in seeking to

differentiate the mutual, formal dependency of

similar figures.

Archaeologists and Art Historians categorize their

objects by creating typologies, thus being able to

make statements about the similarity of objects, about

their purpose, production or style. A computer has no

problem recognizing identically shaped objects, but

has yet to learn our human perception and

understanding of similarity. The approach to this is to

develop shape recognition procedures that link the

degree of simplification and abstraction not only to

human recognition and dissemination patterns as a

means of incrementally evaluating and classifying

unknown objects, but also to categorizations

developed in archaeology and art history. 3D pattern

recognition of the main components must therefore

go hand in hand with archaeological

subcategorization and suitable forms of machine

learning

This paper will show work in progress on

developing those procedures for automatically

generating corpora and will reflect on the associated

schematizations and how they can be applied in

computer science and archaeology. The goal of the

project is to create and evaluate a multi-step

classification process. Eventually, there might be an

object mining that will automatically compare

various grades of similarity and determine to which

category and sub-category (or type) the respective

artifact belongs.

2 RATIONALE AND OBJECTIVE

Classification procedures and pattern recognition

methods are as relevant to visual and object oriented

disciplines as they are to computer science. Both seek

to determine how closely two objects resemble each

other and both use this information for a

classification, even though their objectives differ.

Whereas the goal in computer science is to automate

the classification of unknown objects by means of

pattern matching, typologies created in archaeology

serve as a categorisation criterion for sociocultural

910

Zeckey, A. and Langner, M.

SCHEMATA: 3D Classiﬁcation and Categorization of Ancient Terracotta Figurines.

DOI: 10.5220/0009167809100917

In Proceedings of the 12th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2020) - Volume 2, pages 910-917

ISBN: 978-989-758-395-7; ISSN: 2184-433X

questions regarding the dating, production, or

functioning of artefacts. But both approaches

combine a formal description of the objects with

analytically interpretive approaches.

2.1 Main Objective

Big Data and Cultural Analytics methods require an

appropriate structuring of data, which has not yet

been sufficiently explored for three-dimensional

objects. The methods of 3D pattern recognition are

usually based on cognitive psychology concepts for

object recognition by David Marr and Irving

Biederman. The shape of an object described

geometrically is divided into geometric primitives

and analysed statistically by parts and part

segmentation. Machine learning algorithms help to

automate this process. However, for the classification

of artefacts these methods can provide only rough

approximations. The highly differentiated methods of

biometric face recognition, for example, do not work

with ancient portraits, because their visible

appearance is rather determined by certain hair

designs as by individual face shapes (Schofield et al.,

2012; Lu et al., 2013). These insights lead to the fact

that based on archaeological standards, a

computational feature extraction actually can only be

conducted manually by qualitative shape comparison.

Nevertheless, this process cannot be used

automatically yet. In addition, in areas where the sum

of individual characteristics is too large, too complex

or too heterogeneous to easily create an appropriate

typology, archaeological methods failed. Therefore,

methods of computational shape recognition might be

helpful to define suitable archaeological categories.

In Archaeology and Art History, typologies are

created to make historical and cultural statements.

These qualitative analyses are based on a scientific

framework of classification criteria that are not

necessarily congruent with the concepts of cognitive

psychology, since human perception is not

anthropologically constant, but relies on certain

viewing habits and varies significantly depending on

period or culture. In cases where a large number of

artefacts has quite a similar shape but differs

significantly in certain details, as in serially produced

terracotta figurines that were reworked subsequently,

the concept of typology has reached its limits (Bell,

1993; Burn, 2012). In terms of perception and value

of the figurines, there are too many different criteria

that might bear a meaning. Only a statistical approach

concerning the main features in combination with

archaeological sources and the intrinsic aesthetic

values (such as colour, execution, or style) may solve

the problem.

On the one hand, the algorithms to be developed

must take into account that pictorial works own a

certain complexity of information. They have to

represent the variety of image immanent features in a

better way than a verbal description can provide and

try to follow a genuinely image-oriented logic of

detection and

development when capturing visual

Figure 1: Different grades of similarity in ancient terracotta figurines (after Jeammet, 2003 no. 118–120. © Museum for Fine

Arts Boston).

SCHEMATA: 3D Classiﬁcation and Categorization of Ancient Terracotta Figurines

911

phenomena. On the other hand, it is the task of

Archaeologists to determine the impressions and

viewing habits of the ancient viewer and classify

them in terms of cultural history. By contextually

analyzing the respective conditions of reception, one

tries to reconstruct how the ancient beholder may

have absorbed and processed the visual impressions.

The results of archaeological research must therefore

be equally incorporated into the digital recording of

the pictorial works.

Consequently, the project aims to combine

qualitative and quantitative classification methods to

revise the typology of artefacts. Here, the methods of

computer science (object recognition, shape

comparison and shape analysis) and archaeology

(typology, “Kopienkritik” and contextual analysis)

should benefit from each other, so as to overcome the

aforementioned shortcomings. The conceptual

development-oriented reflection of the approach,

which combines the use of pattern recognition with a

consistent methodological reflection, goes hand in

hand with media reflective studies. The dissertations

that will be developed in the course of this project aim

to conduct preliminary work for the development of

large technical or mental image corpora. However,

both studies also investigate the capabilities of

computer-aided analysis, the limits of this approach

for addressing internal structures, the possibility of

developing novel analytical methods, and the

implications that this approach will have in general

for future archaeological research.

2.2 3D Shape Analysis of Terracotta

Figurines as a Case Study

Ancient terracottas are particularly well suited for

questions of precise classification. The term

“terracottas” refers to figurines made of fired clay that

are not hand sculpted, but rather produced serially

from moulds (Burn, 2012; Erlich, 2015). With regard

to production, distribution, and usage, the items in

question are therefore ancient handcrafted products

that rank below marble and bronze figures in terms of

quality and uniqueness. But they do have the

advantage of having survived in large quantities and

in a wide variety of shapes.

Ancient terracottas resemble each other to

differing degrees. These degrees of resemblance can

be precisely defined by archaeologists and evaluated

progressively by means of classification procedures

at different levels of precision (Muller, 1997): There

are figures that were produced from the same mould

and therefore exhibit an exact correspondence;

alternatively, there are those that were produced in

new moulds using an already fired figurine (Fig.

1a/b/c). These terracottas differ only in size from the

source object. Also, there are figures taken from the

same mould that nonetheless differ in appearance due

to additions or changes by hand (Fig. 1b/c), as a result

of which they no longer belong to the same type. The

next category of terracottas bear strong resemblances

to each other in terms of posture and how the costume

is draped, yet they stem from different moulds (Fig.

1a/d). And finally, there are terracottas in which the

same figure schema occurs in various free

configurations (Fig. 1e). Admittedly, it is possible to

verify at the craftsmanship level that two terracottas

were produced in the same workshop. But if this is

not the case, there are not yet sufficient suitable

criteria for determining degrees of similarity.

New possibilities for artistic formal analysis and

classification can be realised by combining geometric

analysis and information known to archaeologists,

because the traditional archaeological method of

typology is based only on 2D photographs and

subjective judgment which is not as convincing as a

quantitative analysis with 3D models.

3 STATUS OF INTERNATIONAL

RESEARCH AND DISCUSSION

3.1 Applied Computer Science /

Computational Archaeology

In their manual, Juan A. Barceló and Igor Bogdanovic

provide a detailed outline of the current state of

research and an in-depth analysis of how archaeology

and computer science might influence each other

(Barceló / Bogdanovic, 2015). They, too, draw

attention to the fact that economic mass digitisation

of 3D artefacts still constitutes an unsolved problem.

Though the semantic enrichment of 3D data itself

remains challenging, methods for using the geometry

of the 3D shape for data mining is a lively area of

research (e.g. De Luca et al., 2014; Aggarwal, 2015;

Fouhey / Gupta / Zissermann, 2017). Various

methods for recognising 3D objects have been around

for years: CAD models, data-driven geometric

primitives, surface type classification using the

Gaussian image (Amann, 1990. Taylor / Kleeman,

2006) and digital image comparison (Hueting et al.,

2015). They mainly involve automatically extracting

primitives from range data and referring to known

patterns in order to classify unknown objects. The

shape analysis is usually performed statistically

(Dryden / Mardia, 1998). Statistical values describing

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

912

geometric properties of similar shapes are evaluated

with the principal component analysis (PCA)

(Jolliffe, 2002) to analyse the shape variability. In

addition, partial shape matching methods are widely

used (Funkhouser / Shilane, 2006; Bronstein et al.,

2009). Furthermore, outline comparison of one or

more slices of the 3D model (Tal, 2014), as well as

using image-based 3D reconstruction approaches and

formalised primitives in order to generate a library of

elements through the simple declaration of a sequence

of architectural moldings (De Luca et al., 2014), are

utilized. In general, it is much easier to retrieve the

shape of a concentric solid (Hörr, 2011) than that of a

complex structure; the available methods and

technologies so far do not offer a final solution for the

latter. Actual research topics in content-based 3D

object retrieval address different methods. There is

retrieval and classification on textured 3D models as

well as 3D shape retrieval based on distance scanning.

Methods of shape retrieval on non-rigid and large

scale 3D watertight meshes are used as well. They are

complemented with 3D object retrieval with

multimodal views [see titles in the Eurographics

workshops on 3D object retrieval 2014 and 2015].

These different algorithm-based approaches classify

3D models only in terms of basic instances (such as

woman, dog, cup etc).

Thus far, these methods have rarely been used for

the automated capture of artefacts, though

experiments with curve detection and relief detection

are already approved with archaeological artefacts

(Tal, 2014). There are several reasons for this. Firstly,

there are not enough 3D models of sculptures to test

the feasibility of this procedure on a significant scale.

Secondly, works of art (unlike structural elements or

plants, for example) pose significant challenges for

all types of computer-aided classification due to their

complexity and variability. It is much more difficult

to assign a specific instance to a more general class in

this context, because works of art can differ

significantly from each other in terms of shape, size,

and colour. Therefore, a simple computational shape

comparison for “best fit” was used by archaeologists

to analyse the similarity of two artefacts (e. g.

Beenhouwer, 2008). “Best fit” processes are

established in engineering and similar industries and

there are numerous software solutions. These tests are

qualitative rather than quantitative and were already

used for tolerance-based Pass/Fail shape comparison

of ancient sculpture (e.g. www.digital

sculpture.org/laocoon/index.html; Lu et al., 2013;

Frischer, 2014; Rieke-Zapp / Trinkl / Homer, 2017).

The problem with „best fit“ is that only two models

are ever compared to each other, so that a generally

valid extraction of 3D information to compare the

objects does not take place. It is possible to compare

each object with one aother, but only with

morphologically strongly resembling models a

meaningful result can be obtained.

As a result, it is not enough to dismantle the

models into simple geometric forms. A much more

promising approach is to develop shape recognition

procedures that link the degree of simplification and

abstraction not only to human recognition and

dissemination patterns as a means of incrementally

evaluating and classifying unknown objects, but also

to categorizations developed in archaeology and art

history. 3D pattern recognition of the main

components (shape, size, and colour) must therefore

go hand in hand with archaeological sub

categorisation and suitable forms of machine learning

(Bishop, 2006).

3.2 Computational Science

For many years, shape comparison has been an active

research topic at various institutions. Not lastly,

because the shape of a concentric solid is easier to

compare than a complex structure. There is no

definitive solution for these very structures. For this

reason, various basic methods for content-based

shape recognition and shape comparison already exist

in the 2D and 3D area.

However, the often-used partial decomposition

into the basic geometric forms is not suitable for a

final determination of the similarity of complex

works of art, which represent a great challenge due to

their high variability, since much information is lost

and the result becomes too inaccurate. Usually,

similarity is based on a list of numeric attributes

(interest attributes) to be determined. If more than one

feature is specified for matching input features,

similarity is based on averages for each of the interest

attributes. Still, this approach treats all interest

attributes the same and does not match the perception

of the terracotta figurines. Therefore, one must try to

determine a weighting of interest attributes that is

close to both the ancient and modern perception of the

figurines. Here, only archaeological research can help

to distinguish certain types and motives. In addition,

it is more difficult to assign a specific instance to a

general class in this object area, since the differences

SCHEMATA: 3D Classiﬁcation and Categorization of Ancient Terracotta Figurines

913

Figure 2: a - Voxelised model, b - extracted Skeleton Graph, c - Feature Point Extraction of an ancient terracotta figurine

(Göttingen TK23).

shape, size and colour can be considerable. In order

to obtain a more precise result for works of art, pattern

recognition methods mustlink the degree of

simplification and abstraction to human recognition

parameters and in this way redefine a similarity of the

objects. Methods that are based on simple geometric

shapes must not take up the main part of the

recognition process, but rather represent a possible

assistance.

The aim of the project is to develop a possibility

for the automated processing of 2D and 3D data that

goes beyond the usual similarity parameters of

linguistic usage by overcoming everyday paradigms

on the subject of similarity. For this purpose, it is

urgently necessary to shed light on the technical side

of shape comparison and analysis and to evaluate and

combine different methods. As a result, variants of

similarity shall be found that would not be detectable

without the help of the computer, but which have to

be compared with conventional interpretations based

on the parameters of archaeological perception in

order to obtain a useful result. It is important to

consider the differentiation between both sides and to

create a link between the two approaches to the

object. Surely, there is the informatics aspect in which

the aim is to capture and process the complexity of

this data in its entirety in order to deliver new results.

Nevertheless, these results can only be used by the

humanities if they can be combined with established

definitions of archaeological findings or are able to

challenge them. This is why the side of informatics

has to be in constant check with the side of

humanities.

3.3 Planned Implementation

The data to be used will be recorded in the first year

of the project. For this, a number of museums will be

visited to scan nearly 200 Objects. There will be high-

resolution 3D scans of ancient terracottas which will

be created with the structured light scanner in our 3D

Lab. The figurines to be scanned are chosen on the

basis of archaeological terms of similarity.

The first step in the process for recognising

patterns in 3D data should be to carry out a series of

tests with tolerance-based Pass/Fail shape

comparison (“Best Fit”; Figure 3) of the figurines

before moving to shape analysis. During this

procedure, the results will be contrasted with

archaeological theory concerning the concepts of

similarity, seriality, typology and copy. This will

keep the archaeologists and computer scientists in

constant communication and will help to reset

expectations and to define initial criteria of similarity.

The second step in the process for recognising

patterns in 3D data, by means of the description and

segmentation of surface areas (“feature patch”), is to

evaluate the existing matching procedures (statistical

analysis, CAD comparison, structural pattern

recognition) with a view not only to conducting

(semi-)automated processing of large quantities of

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

914

Figure 3: “Best Fit” comparison between the two Erotes

Göttingen TK22 and TK23.

data, but also in terms of quality and correctness. The

goal should be to determine the degree of similarity,

so that the type of resemblance can be determined

from it. To accomplish this, the results of learning

algorithms will have to be constantly compared to

archaeologist’s expectations as a means of identifying

and eliminating system errors in classification (Hörr

et al., 2014).

The next step is to search the 3D models or

segments for common patterns on the basis of the

defined model group. The objective here is to

transcend the tolerance-based shape comparison with

identified methods of shape recognition creating a

process for model-based shape analysis. For these,

algorithms will be tested and validated which have

not been used for this kind of data yet, but seem useful

for a similarity comparison. A rough allocation to the

species and an exact similarity comparison within the

species must be carried out in parallel. Those parts

will converge during the project and go hand in hand

at the end.

The approach might therefore make it possible to

carry out a “best fit” shape comparison with selected

comparative pieces first and then use this comparison

to fine tune the pattern recognition function

progressively from “pose schema” and “figure type”

to “mould identity” (and vice versa). For the goal is

to develop a case study to achieve a finely tuned

categorisation and classification method that goes

beyond verbal, descriptive approaches.

The methods used in this step of development are

various shape recognition techniques of shape

analysis in the 2/3D range, not only to link the degree

of simplification and abstraction with human patterns

of recognition and dissemination for the gradual

evaluation and classification of objects, but to use the

categorisations developed in archaeology and art

history as well. In this regard, extracting feature

regions that distinguish subcategories from each other

and subspace clustering deem to be useful. For this,

Feature Detection and Extraction (Figure 2 c), Image

Labeler, volume-based investigations and

Voxelisation/Skeletonisation (Figure 2 a/b) among

others are used.

After the evaluation of the methods, they are

combined and assigned to a ranking list according to

their stability which is accompanied by an internal

evaluation system when extracting data from an

object.

Thus, although the totality of data that can be

extracted is to be collected, its interpretation is to be

restricted according to parameters that have been

optimised by investigation in order to avoid a

threshold value for defining similarity that is too high

or too low.

Figure 4: Model of the data pipeline.

SCHEMATA: 3D Classiﬁcation and Categorization of Ancient Terracotta Figurines

915

Since the material in the image and object area is

very complex, a partial objective is to test and

compare the different procedures not only for their

productivity, but also for their stability and

effectiveness.

A data pipeline will be experimentally developed for

this purpose. This is an established method in data

analysis that is also suitable for processing big data.

A series of processing elements is connected in a

chain, whereby each step generates the output for the

next step from an input (see Fig. 4 for a general

concept). It implements the different parts of data

processing pipelines that are needed to create

consumable data products: Pre-Processing,

Computing and Post-Processing. With this, it should

be possible to automatically extract data for the

determination and categorisation of similarity in art

historical objects and to create a repository from it.

This repository contains 2D and 3D objects that have

been combined with data that was extracted using

shape analysis. This data can be used for finding new

categorisations or to be linked with existing

humanities categorisations as additional digital

investigations.

4 CONCLUSION

Archaeology as a scientific discipline sees its task

above all in extracting patterns from the sum of

surviving remains of past societies which allows

conclusions to be drawn about the conditions at that

time. For this reason, it has always used forms of

pattern recognition to describe artefacts and images,

although it has continually referred to it more as

structural analysis, typology or seriation. The

question arises whether the methods of

archaeological “Formanalyse” are congruent with the

corresponding methods of digital pattern recognition.

Therefore, the methods will be compared during an

intensive discussion on archaeological concepts for

describing similarity and machine learning

techniques for classification. The discussion has two

objectives: The first is to provide archaeology with

nonverbal forms of description that make it possible

to classify not only typological dependence relations,

but also other degrees of similarity. This may enable

scholars to obtain a more explicit view of the ancient

perception of terracottas concerning types, variants

and motives. The second is to significantly improve

the object mining process, so that a large percentage

of data on objects in a collection can be automatically

stored in databases in the future. On the one hand, this

will revitalise the somewhat deadlocked debate on

types and schemas through the adaptation of

established shape recognition methods from the fields

of mathematics and computer science. On the other

hand, concepts of comparative visual analysis,

developed in visual disciplines, will be applied in the

area of shape recognition. This project will therefore

investigate theoretical aspects of practical

importance, such as a modified definition of the

similarity concept. What does it mean for two shapes

to be similar? How do you describe and define the

uncertainty of the concept? What further conclusions

can be drawn from this for scientific work in

archaeology and computer science?

The capture, analysis and publication of

historically relevant objects as 3D models offers art

historical and archaeological disciplines numerous

advantages: In addition to global availability, simple

and non-intrusive handling, and unlimited

reproducibility, the main advantage is that the

viewpoint is highly adjustable (for example via

rotation, zooming or juxtaposition of objects) as

compared to established documentation methods

(such as orthophotography or plaster casts), thereby

making the objects far more accessible to researchers.

This approach also allows researchers to recreate

historical conditions (in the sense of an object

biography), assign fragments and reconstruct

positioning. As a result, traditional academic

viewpoints and analytical methods will not only be

expanded, but even called into question. The large-

scale virtualisation of objects in collections will in

general have major ramifications for visual

identification processes in historical and visual

disciplines. Also, comparative visualisation of

similarity makes the results of formal analysis

measurable and hence objectifiable, therefore the

visual identification methods used by researchers in

visual disciplines must adapt to new forms of

visualisation which will lead to standardization

processes based on new methodologies. Beginning

with the methodology comparisons proposed in the

project, it will be investigated and described how

archaeological research will be transformed using 3D

models.

REFERENCES

Aggarwal, C. C. 2015. Data Mining: The Textbook.

Heidelberg: Springer.

Amann, H. P. 1990. 3D object recognition based on surface

representations. Neuchâtel.

Barceló, J. A. / I. Bogdanovic (eds.). 2015. Mathematics

and Archaeology. CRC Press, Taylor & Francis Group.

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

916

Beenhouwer, J. de. 2008. Datamanagement for moulded

ceramics and digital image comparison. A case study of

Roman terra cotta figurines. In: Layers of Perception.

Proceedings of the 35th International Conference on

Computer Applications and Quantitative Methods in

Archaeology (CAA), Berlin, Germany, April 2-6, 2007.

Bonn: 160–163.

Bell III, M. 1993. Tanagras and the idea of type, Bulletin

of the Harvard Museum 1: 39–53.

Bishop, C. M. 2006. Pattern Recognition and Machine

Learning. Berlin: Springer.

Bronstein, A. M. / M. M. Bronstein / A. M. Bruckstein / R.

Kimmel. 2009. Partial similarity of objects, or how to

compare a centaur to a horse. Int. J. Comput. Vision 84

no. 2: 163–183.

Burn, L. 2012. Terracottas. In: T. J. Smith / D. Plantzos

(eds.). A companion to Greek art. Chichester: Wiley-

Blackwell: 221–234.

De Luca, L. / D. Lo Buglio. 2014. Geometry vs Semantics:

Open Issues on 3D Reconstruction of Architectural

Elements In: Ioannides M. / E. Quak (eds.). 2014. 3D

Research Challenges in Cultural Heritage. A Roadmap

in Digital Heritage Preservation. Heidelberg / New

York: Springer: 36–49.

Dryden, I. L. / K. V. Mardia. 1998. Statistical Shape

Analysis. New York: John Wiley & Sons.

Erlich, A. 2015. Terracottas. In: E. A. Friedland / M. G.

Sobocinski with E. K. Gazda (eds.), The Oxford

handbook of Roman sculpture. Oxford: Oxford

University Press: 155–172.

Fouhey, D. F. / A. Gupta / A. Zisserman. 2017.

Understanding higher-order shape via 3D shape

attributes. IEEE Transactions on Pattern Analysis and

Machine Intelligence 7 issue 1.

Frischer, B. 2014. 3D data capture, restoration and online

publication of sculpture. In: F. Remondino / S.

Campana (eds.). 3D Recording and Modelling in

Archaeology and Cultural Heritage. Oxford:

Archaeopress: 137–144.

Funkhouser, T. / P. Shilane. 2006. Partial matching of 3d

shapes with priority-driven search. In: Proceedings of

the fourth Eurographics symposium on Geometry

processing, ser. SGP ’06: 131–142.

Hörr, C. / E. Lindinger / G. Brunnett. 2014. Machine

learning based typology development in archaeology.

Journal on Computing and Cultural Heritage (JOCCH)

7.1: 2.1–2.23.

Hörr, C. 2011. Algorithmen zur automatisierten

Dokumentation und Klassifikation archäologischer

Gefäße (Chemnitz 2011).

Hueting, M., M. Ovsjanikov and N. J. Mitra. 2015.

CrossLink: joint understanding of image and 3D model

collections through shape and camera pose variations.

ACM Transactions on Graphics (TOG), 34 (6), 233.

Jeammet, V. (ed.). 2003. Tanagra. Mythe et archéologie.

Exhibition catalogue Musée du Louvre 15 september

2003 – 5 january 2004. Paris: Editions de la Réunion

des musées nationaux.

Jolliffe, I. 2002. Principal Component Analysis, 2nd ed.

Heidelberg: Springer.

Lu, M., Y. Zhang / B. Zheng et al. 2013. Portrait sculptures

of Augustus: Categorization via local shape

comparison. In: 2013 Digital Heritage International

Congress.

Muller, A. 1997. Description et analyse des productions

moulées. Proposition de lexique multilingue,

suggestions de méthode. In: Le moulage en terre cuite

dans l'antiquité. Création et production dérivée,

fabrication et diffusion. Actes du XVIIIe Colloque du

Centre de recherches archéologiques, Lille III, 7 - 8

décembre 1995. Villeneuve d'Ascq: Presses

universitaires du Septentrion: 437–463.

Rieke-Zapp, D / Trinkl, E. / Homer, L. 2017. Face to face

– Considering the moulding of Attic head vases

reconsidering Beazley's groups by quantitative

analysis. Journal of Archaeological Science: Reports

[https://www.sciencedirect.com/science/article/pii/S23

52409X17305345]

Rumscheid, F. 2006. Die figürlichen Terrakotten von

Priene. Wiesbaden: Reichert.

Schofield, D. / K. Lorenz / St. Davy-Jow / M. Anderson.

2012. Roman Portraiture and Biometric Identification.

In: Electronic Visualisation and the Arts (EVA)

London: 163–171 (http://ewic. bcs.

org/content/ConWebDoc/46119).

Tal, A. 2014. 3D Shape Analysis for Archaeology. In:

Ioannides M. / E. Quak (eds.). 3D Research Challenges

in Cultural Heritage. A Roadmap in Digital Heritage

Preservation. Heidelberg / New York: Springer: 50–63.

Taylor, G. / L. Kleeman, 2006. Visual Perception and

Robotic Manipulation: 3D Object Recognition,

Tracking and Hand-Eye Coordination. Heidelberg:

Springer.

SCHEMATA: 3D Classiﬁcation and Categorization of Ancient Terracotta Figurines

917