 Good afternoon, everyone. My name is Alex Netsekal and I'm a PhD student in computer science at the University of Göttingen and I'm working at the Institute for Digital Humanities. Welcome to my short talk about image and object recognition, where I'm going to present our current work and combined with that two upcoming third-party funded projects in our institute. Unfortunately, Professor Lander is not able to attend and sends apologies, so because of that I will present a short summary for both projects myself, so I will have to be sufficient. I'm going to talk about both our projects, Iqrafsun and Schemata, begin by clarifying the aims and techniques, continue with recounting what work has already been done and conclude by giving you an outlook on further possibilities and those respective fields. The project Iqrafsun is a cooperation between the Institute for Digital Humanities of the University of Göttingen under the direction of Professor Martin Langer and the Information Systems and Machine Learning Lab at the University of Hilversheim under the direction of Professor Larsen Thiemel. The project's framework was three years and it will cover two PhD positions and will start this December. Like the other project Schemata, Iqrafsun will also have one PhD student from the humanities and one from the field of computer science to work on each of their fields in a joint project. Basically, the question regarding the identity and overall of the producers of painted crequises has been addressed by archaeology for over 150 years. The computer-assisted analysis of attic wastes is offering the opportunity to put controversial working methods of classical archaeology which were previously based on the expert eye of connoisseurs to the test. For not isolated criteria, but characteristic combinations of details will help objectifying the allocations of painters which can then be compiled and weighted much more easily by computational methods. Additionally, the significance of the archaeology circular method is to be debated controversially because the connoisseurship of individual researchers can hardly be objectified as a hermeneutic basis. Therefore, in the project Iqrafsun, the assignment of wastes to painters, workshops and groups is no longer exemplary and intuitive but on a broad basis of actual data. The aim is to investigate how the attribution criteria are systemized, how they evaluate their relevance and how the importance of similarity networks is historically weighted by precisely naming criteria and arguments. For this purpose, Iqrafsun will develop a data-driven telemetry for attic wastes based on multimodal representations like images and 2D ceramic profiles. In the first step, a selection of waste paintings are being annotated by human experts to mark figures, objects and ornaments. Based on these annotations, we are training a deep convolutional neural network. In the second step, we will determine deep convolutional neural network architectures which can predict whether two waste images of human experts have been assigned to the same painter. From the point of view of machine learning, we have to solve a supervised clustering recall linkage problem for images. We will specifically examine those models that factor to a high degree through semantic representations such as scene descriptions and object glass specific prototypes that allow us to provide clear explanations for the model's decisions. Similar questions about the protest assignment as well as further questions about typical cooperation scenarios between different painters or protestant painters can be answered by exon-mining explanations and outliers of such models. Since the body of attic waste painting is limited and currently compromises about 19,000 vessels, a multi-task learning approach will be used. In cooperation with the Corpus Versorum Antiquovan Germany, a corpus of initially about 20,000 wastes or approximately 50,000 photos will be created. For this, we are using the CVI volumes, 20 painting monographs and digital photos of the museums. Additional tasks such as classifying and describing the objects will enable us to learn better latent representations for waste painting. For the question regarding the relevance of the similarity networks for concrete relationships between the wastes, the use of computers is able to provide impulses because it is possible to describe the type and degree of similarity in a comprehensible way. Another goal is to test digital image classification methods by fundamentally investigating the relationships between archaeological heminoitics, intuitive connoisseurship and data-based objectification and cognition in a central and intensively researched area of classical archaeology. Coming to the second project, the project Chemata, which is done in cooperation with the Gesellschaft für Wissenschaftliche Datenverarbeit in Wöttingen, we are moving into the field of object recognition in terms of working with shape analysis on captured 3D objects of Hellenistic terracottas. The goal of Chemata 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. Based on 200 quite similar terracottas of the late 4th and 3rd centuries BC, a classification system which is able to meet the complexity of the artifacts with digital methods will be elaborated. For this purpose, methods of object mining and 3D data that support the search for a suitable classification and categorization of the images are to be developed. In close cooperation between computer science and archaeology, this experimental process thus 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 by Lucie Böckler and myself. The creation of the corpus for Chemata consisting of 3D models of the mentioned Hellenistic terracottas has already started and has to be in close cooperation with the museums. First scans were taken from the original collection of the archaeological institute in Göttingen as well as from the Martin von Wagner Museum in Wolfsburg. A structured light scanner is being used for achieving the best surface resolution possible using current methods. The data has then progressed, triangulated and pre-processed to be ready for analysis. The pre-processing of the 3D scans is aiming to prepare the data for later processing. For this purpose, the 3D objects are converted to polygithrons, scaled, cleaned, aligned and converted into additional 2D data sets using unwrapping as well as multi-view convolutional neural networks where the output of multiple convolutional neural networks is given as one global descriptor. A different form of pre-processing is required for each calculation performed. Each descriptor characterizes an aspect of the model like geometry or semantics to then be used by dissimilarity functions to measure the distances between the descriptors. Finally, different algorithms can use both of them to classify the data. A global descriptor depicts the whole object with a single vector while local descriptors describe the shape of chosen regions around specific feature points. Both kinds of descriptors are important in 3D computer version and are being used in this project. Global descriptors are being utilized for various possibilities of detecting similarity in two or more objects. Therefore, the object can, for example, be converted to a spherical function or shape functions based on the center or random surface points. This can be used to compute several features of the 3D object automatically. On this slide, you can see three different approaches as a first example that work with random points on the surface of the 3D model where the angles, distances, and the geodesic distance can be measured. A second example shows the orientation and center of the object as determined to then be used with distance measurements. In the last example, you can see the spherical approach dividing the object into sectors along the spherical angles, distributing the model area to construct the shape histogram by counting the number of surface points in each spherical cell. The already mentioned multi-view convolutional neural networks are being used to get virtual cameras around the object, thus taking 2D data into consideration to form ways to progress 3D data with the same possibilities as for 2D data. This complements the 2D investigations also taking place with other 2D data of the real-world objects. Finally, the stances of the objects are analyzed using the scalatization of 3D objects to work with a minimized data set of posture key points from the extracted skeleton. This way, the objects can be compared on a semantic basis to complement the other descriptions of similarity. Local descriptors are also an essential part in this field and already being used in the recording of 3D objects in terms of points to match the single 3D scans taken from every angle of the real-world object. Once those key points are found in the final 3D object, they can be used for a variety of methods based on curvatures, histograms, or back-of-feature techniques, among many others. The challenge in this data is not only the amount of information but also the demands in terms of precision in the humanities. In order to live up to that, a combination of different techniques in the field of local descriptors will be used to complement the already worked-on global descriptors for a better result in similarity. Complementing the field of shape analysis is also best fit from the field of shape comparison, where a tolerance-based approach is used to compare two objects in terms of resemblance. The differences are then given in terms of distances between both measures after they are aligned on the basis of geometrical consistency. This data will be used as a first way to filter through the objects and to then use the more complex calculations of shape analysis on them. To accomplish this, a data pipeline with multiple analyzing methods will be used to determine the computed similarity. Finally, there might be an object mining that will automatically compare various grades of similarity and determine which category and subcategory or type the respective artifact belongs in. In constant balance between archeological and informatic analysis, new criteria of similarity will be defined and evaluated in both disciplines. The approach is therefore making it possible to firstly carry out a best fit shape comparison with selected comparative pieces and secondly use this comparison to fine-tune the pattern recognition function progressively from post-schema and figure type to mode identity. The goal is to develop a case study to achieve a finely tuned categorization and classification method whose capabilities extend beyond those of verbal descriptive approaches. As mentioned before, the pre-processing of the 3D scans aims to prepare the data for later processing. The computing step defines the core of the workflow in which the mentioned methods and selected algorithms are executed with various parameters on the prepared actions. After the data has been obtained from the calculations, the processing and visualization must be taken over in post-processing. Regarding the quality of the results, an algorithm providing the similarity restart can be weighted differently so the outcome of all methods combined will ensure the best result possible. A Grausen will focus completely on 2D images, while Schemata will focus on 3D but also taking 2D analysis into account, each of them aiming to grasp the concept of similarity. Both will work in the restrictive fields to open up new ways of automated similarity detection to make it possible to link similar data to each other automatically without any need for manual annotation. Therefore, artificial intelligence can help to build up large corpora if scientists like archaeology will be able to provide a sufficient number of objects. A multitude of opportunities arise out of this, among which I would like to mention just a few. Automatic recognition of certain image types and motifs, as well as the assignment of fragments, are possible studies for both projects in the future. Also, multimodal research, as well as combining written descriptions with automated detection or in creating written descriptions in an automated way and uniform vocabulary, could be another step. The combination of a first-distant viewing of the material with a follow-up close viewing of the individual objects is present in both objects, as it is a core problem of classification. Lastly, the connoissorship problem remains an active field of research and it has to be ex-amined how far the automatic digital classification is the same as the connoissar knowledge and if not in which ways it might complement given structures. Thank you for your time.