This action involves the management of the project’s work packages, deliverables and resources. Throughout the project, all work packages and deliverables will be monitored to ensure their efficient completion within their prescribed time frames and budget. To ensure this, frequent communications will take place between team members, both physically and over video conference calls, to warranty everyone is up to date and all issues are tackled in time.
This action includes all dissemination and communication activities for spreading the research ideas, results and deliverables to the research community and the public in Cyprus and abroad. The involved dissemination activities are 2-3 journal publications and 3-4 conference publications, along with communication activities (project website, YouTube channel and participation/organisation of various events). The attended/organised events include the Excellence Hubs Open Day, the Computer Science Day and the RIF’s Researcher’s Night.
Dataset and User Requirements Collection
This action involves the dataset and user requirement collection. First, a user study was conducted with the participation of cultural heritage experts (e.g., architects, academics, historians, etc.), the results of which were used to define their requirements regarding cultural heritage monuments’ documentation. These were useful both to understand how to collect the necessary data and how to structure the monument analysis platform. After this, a list of candidate monuments was compiled in coordination with experts. The list covers buildings from different time periods (e.g., Byzantine, Venetian, Ottoman, etc.) that are located in Cyprus. The most promising were collected in 3D form (i.e. mesh), either from already existing 3D repositories or structure-from-motion photogrammetry. The final dataset contains 30 monuments located in Cyprus’ capital, and are stylistically (e.g. Gothic, Byzantine) and structurally labelled (e.g. wall, column, arch). The dataset was used for the training of the Structure and Style ANNs and the monuments are publicly available to be viewed in the project’s platform.
Structure ANN Development
This action involves basic research. First, a way to convert each 3D monument mesh into a suitable representation to feed to the ANN input layer was chosen. Two competing shape representations, based on randomly sampled point clouds or volumetric occupancy grid were investigated. The point-based method that was test was the popular PoineNet++ and for the volumetric case, the arising O-CNN. Specifically, the volumetric implementation of High Resolution Network (HRNet) using the convolution operations proposed by the authors of O-CNN. A set of classic features, suitable for distinguishing shape parts (e.g. normal vector, colour) were chosen, based on consultation of the literature in computer graphics. Both networks were trained using a transfer learning technique, that is they were initially trained on a larger dataset (BuildingNet), to learn building features, and then fine-tuned on the Cypriot dataset, to be specialised on monuments. Their performance was evaluated through standard classification accuracy metrics and popular semantic segmentation metrics (e.g., Intesection over Union). The best performing representation and formulation was chosen and integrated with the platform, which was the volumetric implementation of the HRNet (O-HRNet).
Style ANN Development
This action also involves basic research. The methodology is almost identical to the one used for the development of the Structure ANN. The difference lies in the type of input representation, neural network model and features used. The literature in architecture as well as archaeology will be consulted in order to incorporate any empirical observations of what characterizes style into the proposed ANN.
Inverse Procedural Modeling and Procedural Analysis
This action involves basic research. First, the literature in inverse procedural modeling was reviewed to identify the state-of-the-art methods. This assisted in the formulation and development of the shape grammar extraction and comparison method. Following this review, a method was designed and developed to extract a procedural ruleset (shape grammar) based on the structurally segmented monument. Last, a method was designed and developed for comparing two procedural rulesets obtained from the extraction method, to identify and reveal any similarities or differences in the design rules for two monuments. This comparison is far from trivial due to construction ambiguities, but given the fact that both grammars are extracted by the same method they are directly comparable. The grammar extraction and the grammar comparison methods were to be used in the shape-grammar-based design and comparison tool of the project’s platform, but because of its dependency to other frameworks (e.g., Houdini) it could not be integrated. This component can be used individually be following the instructions on the code site.
Code can be found here: https://github.com/dagush/ANNFASS-Girona
Monument Analysis Software Package Development
This action involves applied research and technology development. Given the user requirements collected from the cultural heritage experts a platform was designed, using micro-services, to accommodate their needs. Specifically, the tool allows for the storing, sharing, viewing and comparing of monuments, along with their structural and stylistic labels. Each monument has a dedicated page that contains a plethora of information related to it (e.g. description, architect) to help experts gain a more complete image of the monument. New monuments can be added to the platform and using the integrated Structure and Style ANNs acquire automatically their structure and style labels.
To try the platform visit: https://annfass-srv.cs.ucy.ac.cy/home